Author Archives: Bart Thomee

Report from the MMM 2020 Special Session on Multimedia Datasets for Repeatable Experimentation (MDRE 2020)

By | | , , ,
Leave a comment

Introduction

Information retrieval and multimedia content access have a long history of comparative evaluation, and many of the advances in the area over the past decade can be attributed to the availability of open datasets that support comparative and repeatable experimentation. Hence, sharing data and code to allow other researchers to replicate research results is needed in the multimedia modeling field, as it helps to improve the performance of systems and the reproducibility of published papers.

This report summarizes the special session on Multimedia Datasets for Repeatable Experimentation (MDRE 2020), which was organized at the 26th International Conference on MultiMedia Modeling (MMM 2020), held in January 2020 in Daejeon, South Korea.

The intent of these special sessions is to be a venue for releasing datasets to the multimedia community and discussing dataset related issues. The presentation mode in 2020 was to have short presentations (approximately 8 minutes), followed by a panel discussion moderated by Aaron Duane. In the following we summarize the special session, including its talks, questions, and discussions.

Presentations

GLENDA: Gynecologic Laparoscopy Endometriosis Dataset

The session began with a presentation on ‘GLENDA: Gynecologic Laparoscopy Endometriosis Dataset’ [1], given by Andreas Leibetseder from the University of Klagenfurt. The researchers worked with experts on gynecologic laparoscopy, a type of minimally invasive surgery (MIS), that is performed via a live feed of a patient’s abdomen to survey the insertion and handling of various instruments for conducting medical treatments. Adopting this kind of surgical intervention not only facilitates a great variety of treatments but also the possibility of recording such video streams is essential for numerous post-surgical activities, such as treatment planning, case documentation and education. The process of manually analyzing these surgical recordings, as it is carried out in current practice, usually proves tediously time-consuming. In order to improve upon this situation, more sophisticated computer vision as well as machine learning approaches are actively being developed. Since most of these approaches rely heavily on sample data that, especially in the medical field, is only sparsely available, the researchers published the Gynecologic Laparoscopy ENdometriosis DAtaset (GLENDA) – an image dataset containing region-based annotations of a common medical condition called endometriosis. 

Endometriosis is a disorder involving the dislocation of uterine-like tissue. Andreas explained that this dataset is the first of its kind and was created in collaboration with leading medical experts in the field. GLENDA contains over 25K images, about half of which are pathological, i.e., showing endometriosis, and the other half non-pathological, i.e., containing no visible endometriosis. The accompanying paper thoroughly described the data collection process, the dataset’s properties and structure, while also discussing its limitations. The authors plan on continuously extending GLENDA, including the addition of other relevant categories and ultimately lesion severities. Furthermore, they are in the process of collecting specific ”endometriosis suspicion” class annotations in all categories for capturing a common situation where at times it proves difficult, even for endometriosis specialists, to classify the anomaly without further inspection. The difficulty in classification may be due to several reasons, such as visible video artifacts. Including such challenging examples in the dataset may greatly improve the quality of endometriosis classifiers.

Kvasir-SEG: A Segmented Polyp Dataset

The second presentation was given by Debesh Jha from the Simula Research Laboratory, who introduced the work entitled ‘Kvasir-SEG: A Segmented Polyp Dataset’ [2]. Debesh explained that pixel-wise image segmentation is a highly demanding task in medical image analysis. Similar to the aforementioned GLENDA dataset, it is difficult to find annotated medical images with corresponding segmentation masks in practice. The Kvasir-SEG dataset is an open-access corpus of gastrointestinal polyp images and corresponding segmentation masks, which has been further manually annotated and verified by an experienced gastroenterologist. The researchers demonstrated the use of their dataset with both a traditional segmentation approach and a modern deep learning-based CNN approach. In addition to presenting the Kvasir-SEG dataset, Debesh also discussed the FCM clustering algorithm and the ResUNet-based approach for automatic polyp segmentation they presented in their paper. The results show that the ResUNet model was superior to FCM clustering.

The researchers released the Kvasir-SEG dataset as an open-source dataset to the multimedia and medical research communities, in the hope that it can help evaluate and compare existing and future computer vision methods. By adding segmentation masks to the Kvasir dataset, which until today only consisted of framewise annotations, the authors have enabled multimedia and computer vision researchers to contribute in the field of polyp segmentation and automatic analysis of colonoscopy videos. This could boost the performance of other computer vision methods and may be an important step towards building clinically acceptable CAI methods for improved patient care.

Rethinking the Test Collection Methodology for Personal Self-Tracking Data

The third presentation was given by Cathal Gurrin from Dublin City University and was titled ‘Rethinking the Test Collection Methodology for Personal Self-Tracking Data’ [3]. Cathal argued that, although vast volumes of personal data are being gathered daily by individuals, the MMM community has not really been tackling the challenge of developing novel retrieval algorithms for this data, due to the challenges of getting access to the data in the first place. While initial efforts have taken place on a small scale, it is their conjecture that a new evaluation paradigm is required in order to make progress in analysing, modeling and retrieving from personal data archives. In their position paper, the researchers proposed a new model of Evaluation-as-a-Service that re-imagines the test collection methodology for personal multimedia data in order to address the many challenges of releasing test collections of personal multimedia data. 

After providing a detailed overview of prior research on the creation and use of self-tracking data for research, the authors identified issues that emerge when creating test collections of self-tracking data as commonly used by shared evaluation campaigns. This includes in particular the challenge of finding self-trackers willing to share their data, legal constraints that require expensive data preparation and cleaning before a potential release to the public, as well as ethical considerations. The Evaluation-as-a-Service model is a novel evaluation paradigm meant to address these challenges by enabling collaborative research on personal self-tracking data. The model relies on the idea of a central data infrastructure that guarantees full protection of the data, while at the same time allowing algorithms to operate on this protected data. Cathal highlighted the importance of data banks in this scenario. Finally, he briefly outlined technical aspects that would allow setting up a shared evaluation campaign on self-tracking data.

Experiences and Insights from the Collection of a Novel Multimedia EEG Dataset

The final presentation of the session was also provided by Cathal Gurrin from Dublin City University in which he introduced the topic ‘Experiences and Insights from the Collection of a Novel Multimedia EEG Dataset’ [4]. This work described how there is a growing interest in utilising novel signal sources such as EEG (Electroencephalography) in multimedia research. When using such signals, subtle limitations are often not readily apparent without significant domain expertise. Multimedia research outputs incorporating EEG signals can fail to be replicated when only minor modifications have been made to an experiment or seemingly unimportant (or unstated) details are changed. Cathal claimed that this can lead to over-optimistic or over-pessimistic viewpoints on the potential real-world utility of these signals in multimedia research activities.

In their paper, the researchers described the EEG/MM dataset and presented a summary of distilled experiences and knowledge gained during the preparation (and utilisation) of the dataset that supported a collaborative neural-image labelling benchmarking task. They stated that the goal of this task was to collaboratively identify machine learning approaches that would support the use of EEG signals in areas such as image labelling and multimedia modeling or retrieval. The researchers stressed that this research is relevant for the multimedia community as it suggests a template experimental paradigm (along with datasets and a baseline system) upon which researchers can explore multimedia image labelling using a brain-computer interface. In addition, the paper provided insights and experience of commonly encountered issues (and useful signals) when conducting research that utilises EEG in multimedia contexts. Finally, this work provided insight on how an EEG dataset can be used to support a collaborative neural-image labelling benchmarking task.

Discussion

After the presentations, Aaron Duane moderated a panel discussion in which all presenters participated, as well as Björn Þór Jónsson who joined the panel as one of the special session chairs.

The panel began with a question about how the research community should address data anonymity in large multimedia datasets and how, even if the dataset is isolated and anonymised, data analysis techniques can be utilised to reverse this process either partially or completely. The panel agreed this was an important question and acknowledged that there is no simple answer. Cathal Gurrin stated that there is less of a restrictive onus on the datasets used for such research because the owners of the dataset often provide it with full knowledge of how it will be used.

As a follow up, the questioner asked the panel about GDPR compliancy in this context and the fact that uploaders could potentially change their minds about allowing their datasets to be used in research several years after it was released. The panel acknowledged this remains an open concern and even expanded on such concerns by presenting an additional concern, namely the malicious uploading of data without the consent of the owner. One solution to this which was provided by the panel was the introduction of an additional layer of security in the form of a human curator who could review the security and privacy concerns of a dataset during its generation, as is the case with some datasets of personal data currently under release to the community. 

The discussion continued with much interest continuing to be directed toward effective privacy in datasets, especially when dealing with personal data, such as those generated by lifeloggers. One audience member recalled a story where a personal dataset was publicly released and individuals were able to garner personal information about individuals who were not the original uploader of the dataset and who did not consent to their face or personal information being publicly released. Cathal and Björn acknowledged that this remains an issue but drew attention to advanced censoring techniques such as automatic face blurring which is rapidly maturing in the domain. Furthermore, they claimed that the proposed model of Evaluation-as-a-Service discussed in Cathal’s earlier presentation could help to further alleviate some of these concerns.

Steering the conversation away from exclusively dealing with data privacy concerns, Aaron directed a question at Debesh and Andreas regarding the challenges and limitations associated with working directly with medical professionals to generate their datasets related to medical disorders. Debesh stated that there were numerous challenges such as the medical professionals being unfamiliar with the tools used in the generation of this work and that in many cases circumstances required multiple medical professionals and their opinion as they would often disagree. This generated significant technical and administrative overhead for the researchers and their work which resulted in a tedious speed of progress. Andreas stated that such issues were identical for him and his colleagues and highlighted the importance of effective communication between the medical experts and the technical researchers.

Towards the end of the discussion, the panel discussed the concept of encouraging the release of more large-scale multimedia datasets for experimentation and what challenges are currently associated with that. The panel responded that the process remains difficult but having special sessions such as this are very helpful. The recognition of papers associated with multimedia datasets is becoming increasingly apparent with many exceptional papers earning hundreds of citations within the community. The panel also stated that we should be mindful of the nature of each dataset as releasing the same type of dataset, again and again, is not beneficial and has the potential to do more harm than good.

Conclusions

The MDRE special session, in its second incarnation at MMM 2020, was organised to facilitate the publication of high-quality datasets, and for community discussions on the methodology of dataset creation. The creation of reliable and shareable research artifacts, such as datasets with reliable ground truths, usually represents tremendous effort; effort that is rarely valued by publication venues, funding agencies or research institutions. In turn, this leads many researchers to focus on short-term research goals, with an emphasis on improving results on existing and often outdated datasets by small margins, rather than boldly venturing where no researchers have gone before. Overall, we believe that more emphasis on reliable and reproducible results would serve our community well, and the MDRE special session is a small effort towards that goal.

Acknowledgements

The session was organized by the authors of the report, in collaboration with Duc-Tien Dang-Nguyen (Dublin City University), who could not attend MMM. The panel format of the special session made the discussions much more engaging than that of a traditional special session. We would like to thank the presenters, and their co-authors for their excellent contributions, as well as the members of the audience who contributed greatly to the session.

References

  • [1] Leibetseder A., Kletz S., Schoeffmann K., Keckstein S., and Keckstein J. “GLENDA: Gynecologic Laparoscopy Endometriosis Dataset.” In: Cheng WH. et al. (eds) MultiMedia Modeling. MMM 2020. Lecture Notes in Computer Science, vol. 11962, 2020. Springer, Cham. https://doi.org/10.1007/978-3-030-37734-2_36.
  • [2] Jha D., Smedsrud P.H., Riegler M.A., Halvorsen P., De Lange T., Johansen D., and Johansen H.D. “Kvasir-SEG: A Segmented Polyp Dataset.” In: Cheng WH. et al. (eds) MultiMedia Modeling. MMM 2020. Lecture Notes in Computer Science, vol. 11962, 2020. Springer, Cham. https://doi.org/10.1007/978-3-030-37734-2_37.
  • [3] Hopfgartner F., Gurrin C., and Joho H. “Rethinking the Test Collection Methodology for Personal Self-tracking Data.” In: Cheng WH. et al. (eds) MultiMedia Modeling. MMM 2020. Lecture Notes in Computer Science, vol. 11962, 2020. Springer, Cham. https://doi.org/10.1007/978-3-030-37734-2_38.
  • [4] Healy G., Wang Z., Ward T., Smeaton A., and Gurrin C. “Experiences and Insights from the Collection of a Novel Multimedia EEG Dataset.” In: Cheng WH. et al. (eds) MultiMedia Modeling. MMM 2020. Lecture Notes in Computer Science, vol. 11962, 2020. Springer, Cham. https://doi.org/10.1007/978-3-030-37734-2_39.

Dataset Column: ToCaDa Dataset with Multi-Viewpoint Synchronized Videos

By | | , ,
Leave a comment

This column describes the release of the Toulouse Campus Surveillance Dataset (ToCaDa). It consists of 25 synchronized videos (with audio) of two scenes recorded from different viewpoints of the campus. An extensive manual annotation comprises all moving objects and their corresponding bounding boxes, as well as audio events. The annotation was performed in order to i) enhance audiovisual objects that can be visible, audible or both, according to each recording location, and ii) uniquely identify all objects in each of the two scenes. All videos have been «anonymized». The dataset is available for download here.

Introduction

The increasing number of recording devices, such as smartphones, has led to an exponential production of audiovisual documents. These documents may correspond to the same scene, for instance an outdoor event filmed from different points of view. Such multi-view scenes contain a lot of information and provide new opportunities for answering high-level automatic queries.

In essence, these documents are multimodal, and their audio and video streams contain different levels of information. For example, the source of a sound may either be visible or not according to the different points of view. This information can be used separately or jointly to achieve different tasks, such as synchronising documents or following the displacement of a person. The analysis of these multi-view field recordings further allows understanding of complex scenarios. The automation of these tasks faces a need for data, as well as a need for the formalisation of multi-source retrieval and multimodal queries. As also stated by Lefter et al., “problems with automatically processing multimodal data start already from the annotation level” [1]. The complexity of the interactions between modalities forced the authors to produce three different types of annotations: audio, video, and multimodal.

In surveillance applications, humans and vehicles are the most important common elements studied. In consequence, detecting and matching a person or a car that appears in several videos is a key problem. Although many algorithms have been introduced, a major relative problem still is how to precisely evaluate and to compare these algorithms in reference to a common ground truth. Datasets are required for evaluating multi-view based methods.

During the last decade, public datasets have become more and more available, helping with the evaluation and comparison of algorithms, and in doing so, contributing to improvements in human and vehicle detection and tracking. However, most of the datasets focus on a specific task and do not support the evaluation of approaches that mix multiple sources of information. Only few datasets provide synchronized videos with overlapping fields of view. Yet, these rarely provide more than 4 different views even though more and more approaches could benefit from having additional views available. Moreover, soundtracks are almost never provided despite being a rich source of information, as voices and motor noises can help to recognize, respectively, a person or a car.

Notable multi-view datasets are the following.

  • The 3D People Surveillance Dataset (3DPeS) [2] comprises 8 cameras with disjoint views and 200 different people. Each person appears, on average, in 2 views. More than 600 video sequences are available. Thus, it is well-suited for people re-identification. Cameras parameters are provided, as well as a coarse 3D reconstruction of the surveilled environment.
  • The Video Image Retrieval and Analysis Tool (VIRAT) [3] dataset provides a large amount of surveillance videos with a high pixel resolution. In this dataset, 16 scenes were recorded for hours although in the end only 25 hours with significant activities were kept. Moreover, only two pairs of videos present overlapping fields of view. Moving objects were annotated by workers with bounding boxes, as well as some buildings or areas. Three types of events were also annotated, namely (i) single person events, (ii) person and vehicle events, and (iii) person and facility events, leading to 23 classes of events. Most actions were performed by people with minimal scripted actions, resulting in realistic scenarios with frequent incidental movers and occlusions.
  • Purely action-oriented datasets can be found in the Multicamera Human Action Video (MuHAVi) [4] dataset, in which 14 actors perform 17 different action classes (such as “kick”, “punch”, “gunshot collapse”) while 8 cameras capture the indoor scene. Likewise, Human3.6M [5] contains videos where 11 actors perform 15 different classes of actions while being filmed by 4 digital cameras; its specificity lies in the fact that 1 time-of-flight sensor and 10 motion cameras were also used to estimate and to provide the 3DT pose of the actors on each frame. Both background subtraction and bounding boxes are provided at each frame. In total, more than 3.6M frames are available. In these two datasets, actions are performed in unrealistic conditions as the actors follow a script consisting of actions that are performed one after the other.

In the table below a comparison is shown between the aforementioned datasets, which are contrasted with the new ToCaDa dataset we recently introduced and describe in more detail below.

Properties 3DPeS [2] VIRAT [3] MuHAVi [4] Human3.6M [5] ToCaDa [6]
# Cameras 8 static 16 static 8 static 4 static 25 static
# Microphones 0 0 0 0 25+2
Overlapping FOV Very partially 2+2 8 4 17
Disjoint FOV 8 12 0 0 4
Synchronized No No Partially Yes Yes
Pixel resolution 704 x 576 1920 x 1080 720 x 576 1000 x 1000 Mostly 1920 x 1080
# Visual objects 200 Hundreds 14 11 30
# Action types 0 23 17 15 0
# Bounding boxes 0 ≈ 1 object/second 0 ≈ 1 object/frame ≈ 1 object/second
In/outdoor Outdoor Outdoor Indoor Indoor Outdoor
With scenario No No Yes Yes Yes
Realistic Yes Yes No No Yes

ToCaDa Dataset

As a large multi-view, multimodal, and realistic video collection does not yet exist, we therefore took the initiative to produce such a dataset. The ToCaDa dataset [6] comprises 25 synchronized videos (including soundtrack) of the same scene recorded from multiple viewpoints. The dataset follows two detailed scenarios consisting of comings and goings of people, cars and motorbikes, with both overlapping and non-overlapping fields of view (see Figures 1-2). This dataset aims at paving the way for multidisciplinary approaches and applications such as 4D-scene reconstruction, object re-identification/tracking and multi-source metadata modeling and querying.

Figure 1: The campus contains 25 cameras, of which 8 are spread out across the area and 17 are located within the red rectangle (see Figure 2).
Figure 2: The main building where 17 cameras with overlapping fields of view are concentrated.

About 20 actors were asked to follow two realistic scenarios by performing scripted actions, like driving a car, walking, entering or leaving a building, or holding an item in hand while being filmed. In addition to ordinary actions, some suspicious behaviors are present. More precisely:

  • In the first scenario, a suspect car (C) with two men inside (D the driver and P the passenger) arrives and parks in front of the main building (within the sights of the cameras with overlapping views). P gets out of the car C and enters the building. Two minutes later, P leaves the building holding a package and gets in C. C leaves the parking (see Figure 3) and gets away from the university campus (passing in front of some of the disjoint fields of view cameras). Other vehicles and persons regularly move in different cameras with no suspicious behavior.
  • In the second scenario, a suspect car (C) with two men inside (D the driver and P the passenger) arrives and parks badly along the road. P gets out of the car and enters the building. Meanwhile, a women W knocks on the car window to ask the driver D to park correctly, but he drives off immediately. A few minutes later, P leaves the building with a package and seems confused as the car is missing. He then runs away. In the end, in one of the disjoint-view cameras, we can see him waiting until C picks him up.
Figure 3: A subset of all the synchronized videos for a particular frame of the first scenario. First row: cameras located in front of the building. Second and third rows: cameras that face the car park. A car is circled in red to highlight the largely overlapping fields of view.

The 25 camera holders we enlisted used their own mobile devices to record the scene, leading to a large variety of resolutions, image quality, frame rates and video duration. Three foghorns were blown in order to coordinate this heterogeneous disposal:

  • The first one stands for a warning 20 seconds before the start, to give enough time to start shooting.
  • The second one is the actual starting time, used to temporally synchronize the videos.
  • The third one indicates the ending time.

All the videos were collected and were manually synchronized using the second and the third foghorn blows as starting and ending times. Indeed, the second one can be heard at the beginning of every video.

Annotations

A special annotation procedure was set to handle the audiovisual content of this multi-view data [7]. Audio and video parts of each document were first separately annotated, after which a fusion of these modalities was realized.

The ground truth annotations are stored in json files. Each file corresponds to a video and shares the same title but not the same extension, namely <video_name>.mp4 annotations are stored in <video_name>.json. Both visual and audio annotations are stored together in the same file.

By annotating, our goal is to detect the visual objects and the salient sound events and, when possible, to associate them. Thus, we have grouped them into the generic term audio-visual object. This way, the appearance of a vehicle and its motor sound will constitute a single coherent audio-visual object and is associated with the same ID. An object that can be seen but cannot be heard is also an audio-visual object but with only a visual component, and similarly for an object that can only be heard. An example is given in Listing 1.

Listing 1: Json file structure of the visual component of an object in a video, visible from 13.8s to 18.2s and from 29.72s to 32.28s and associated with id 11.

To help with the annotation process, we developed a program for navigating through the frames of the synchronized videos and for identifying audio-visual objects by drawing bounding boxes in particular frames and/or specifying starting and ending times of salient sound. Bounding boxes were drawn around every moving object with a flag indicating whether the object was fully visible or occluded, specifying its category (human or vehicle), providing visual details (for example clothes types or colors), and timestamps of its apparitions and disappearances. Audio events were also annotated by a category and two timestamps.

Regarding bounding boxes, the coordinates of top-left and bottom-right corners of the bounding boxes are given. Bounding boxes were drawn such that the object is fully contained within the box and as tight as possible. For this purpose, our annotation tool allows the user to draw an initial approximate bounding box and then to adjust its boundaries at a pixel-level.

As drawing one bounding box for each object on every frame requires a huge amount of time, we have drawn bounding boxes on a subset of frames, so that the intermediate bounding boxes of an object can be linearly interpolated using its previous and next drawn bounding boxes. On average, we have drawn one bounding box per second for humans and two for vehicles due to their speed variation. For objects with irregular speed or trajectory, we have drawn more bounding boxes.

Regarding the audio component of an audio-visual object, namely the salient sound events, an audio category (voice, motor sound) is given in addition to its ID, as well as a list of details and time bounds (see Listing 2).

Listing 2: Json file structure of an audio event in a given video. As it is associated with id 11, it corresponds to the same audio-visual object as the one in Listing 1.

Finally, we linked the audio to the video objects, by giving the same ID to the audio object in case of causal identification, which means that the acoustic source of the audio event is the object (a car or a person for instance) that was annotated. This step was particularly crucial, and could not be automatized, as a complex expertise is required to identify the sound sources. For example, in the video sequence illustrated in Figure 4, a motor sound is audible and seems to come from the car whereas it actually comes from a motorbike behind the camera.

Figure 4: At this time of the video sequence of camera 10, a motor sound is heard and seems to come from the car while it actually comes from a motorbike behind the camera.

In case of an object presenting different sound categories (a car with door slams, music and motor sound for example), one object is created for each category and the same ID is given.

Ethical and Legal

According to the European legislation, it is forbidden to make images publicly available of people who might be recognized or of license plates. As people and license plates are visible in our videos, to conform to the General Data Protection Regulation (GDPR) we decided to:

  • Ask actors to sign an authorization for publishing their image, and
  • Apply post treatment on videos to blur faces of other people and any license plates.

Conclusion

We have introduced a new dataset composed of two sets of 25 synchronized videos of the same scene with 17 overlapping views and 8 disjoint views. Videos are provided with their associated soundtracks. We have annotated the videos by manually drawing bounding boxes on moving objects. We have also manually annotated audio events. Our dataset offers simultaneously a large number of both overlapping and disjoint synchronized views and a realistic environment. It also provides audio tracks with sound events, high pixel resolution and ground truth annotations.

The originality and the richness of this dataset come from the wide diversity of topics it covers and the presence of scripted and non-scripted actions and events. Therefore, our dataset is well suited for numerous pattern recognition applications related to, but not restricted to, the domain of surveillance. We describe below, some multidisciplinary applications that could be evaluated using this dataset:

3D and 4D reconstruction: The multiple cameras sharing overlapping fields of view along with some provided photographs of the scene allow performing a 3D reconstruction of the static parts of the scene and to retrieve intrinsic parameters and poses of the cameras using a Structure-from-Motion algorithm. Beyond a 3D reconstruction, the temporal synchronization of the videos could enable to render dynamic parts of the scene as well and to obtain a 4D reconstruction.

Object recognition and consistent labeling: Evaluation of algorithms for human and vehicle detection and consistent labeling across multiple views can be performed using the annotated bounding boxes and IDs. To this end, overlapping views provide a 3D environment that could help to infer the label of an object in a video knowing its position and label in another video.

Sound event recognition: The audio events recorded from different locations and manually annotated provide opportunities to evaluate the relevance of consistent acoustic models by, for example, launching the identification and indexing of a specific sound event. Looking for a particular sound by similarity is also feasible.

Metadata modeling and querying: The multiple layers of information of this dataset, both low-level (audio/video signal) and high-level (semantic data available in the ground truth files) enable handling of information at different resolutions of space and time, allowing to perform queries on heterogeneous information.

References

[1] I. Lefter, L.J.M. Rothkrantz, G. Burghouts, Z. Yang, P. Wiggers. “Addressing multimodality in overt aggression detection”, in Proceedings of the International Conference on Text, Speech and Dialogue, 2011, pp. 25-32.
[2] D. Baltieri, R. Vezzani, R. Cucchiara. “3DPeS: 3D people dataset for surveillance and forensics”, in Proceedings of the 2011 joint ACM workshop on Human Gesture and Behavior Understanding, 2011, pp. 59-64.
[3] S. Oh, A. Hoogs, A. Perera, N. Cuntoor, C. Chen, J.T. Lee, S. Mukherjee, J.K. Aggarwal, H. Lee, L. Davis, E. Swears, X. Wang, Q. Ji, K. Reddy, M. Shah, C. Vondrick, H. Pirsiavash, D. Ramanan, J. Yuen, A. Torralba, B. Song, A. Fong, A. Roy-Chowdhury, M. Desai. “A large-scale benchmark dataset for event recognition in surveillance video”, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2011, pp. 3153-3160.
[4] S. Singh, S.A. Velastin, H. Ragheb. “MuHAVi: A multicamera human action video dataset for the evaluation of action recognition methods”, in Proceedings of the 7th IEEE International Conference on Advanced Video and Signal Based Surveillance, 2010, pp. 48-55.
[5] C. Ionescu, D. Papava, V. Olaru, C. Sminchisescu. “Human3.6M: Large scale datasets and predictive methods for 3d human sensing in natural environments”, IEEE transactions on Pattern Analysis and Machine Intelligence, 36(7), 2013, pp. 1325-1339.
[6] T. Malon, G. Roman-Jimenez, P. Guyot, S. Chambon, V. Charvillat, A. Crouzil, A. Péninou, J. Pinquier, F. Sèdes, C. Sénac. “Toulouse campus surveillance dataset: scenarios, soundtracks, synchronized videos with overlapping and disjoint views”, in Proceedings of the 9th ACM Multimedia Systems Conference. 2018, pp. 393-398.
[7] P. Guyot, T. Malon, G. Roman-Jimenez, S. Chambon, V. Charvillat, A. Crouzil, A. Péninou, J. Pinquier, F. Sèdes, C. Sénac. “Audiovisual annotation procedure for multi-view field recordings”, in Proceedings of the International Conference on Multimedia Modeling, 2019, pp. 399-410.

Dataset Column: Report from the MMM 2019 Special Session on Multimedia Datasets for Repeatable Experimentation (MDRE 2019)

By | | , ,
Leave a comment

Special Session

Information retrieval and multimedia content access have a long history of comparative evaluation, and many of the advances in the area over the past decade can be attributed to the availability of open datasets that support comparative and repeatable experimentation. Sharing data and code to allow other researchers to replicate research results is needed in the multimedia modeling field, as it helps to improve the performance of systems and the reproducibility of published papers.

This report summarizes the special session on Multimedia Datasets for Repeatable Experimentation (MDRE 2019), which was organized at the 25th International Conference on MultiMedia Modeling (MMM 2019), which was held in January 2019 in Thessaloniki, Greece.

The intent of these special sessions is to be a venue for releasing datasets to the multimedia community and discussing dataset related issues. The presentation mode in 2019 was to have short presentations (8 minutes) with some questions, and an additional panel discussion after all the presentations, which was moderated by Björn Þór Jónsson. In the following we summarize the special session, including its talks, questions, and discussions.

The special session presenters: Luca Rossetto, Cathal Gurrin and Minh-Son Dao.

Presentations

A Test Collection for Interactive Lifelog Retrieval

The session started with a presentation about A Test Collection for Interactive Lifelog Retrieval [1], given by Cathal Gurrin from Dublin City University (Ireland). In their work, the authors introduced a new test collection for interactive lifelog retrieval, which consists of multi-modal data from 27 days, comprising nearly 42 thousand images and other personal data (health and activity data; more specifically, heart rate, galvanic skin response, calorie burn, steps, blood pressure, blood glucose levels, human activity, and diet log). The authors argued that, although other lifelog datasets already exist, their dataset is unique in terms of the multi-modal character, and has a reasonable and easily manageable size of 27 consecutive days. Hence, it can also be used for interactive search and provides newcomers with an easy entry into the field. The published dataset has already been used for the Lifelog Search Challenge (LSC) [5] in 2018, which is an annual competition run at the ACM International Conference on Multimedia Retrieval (ICMR).

The discussion about this work started with a question about the plans for the dataset and whether it should be extended over the years, e.g. to increase the challenge of participating in the LSC. However, the problem with public lifelog datasets is the fact that there is a conflict between releasing more content and safeguarding privacy. There is a strong need to anonymize the contained images (e.g. blurring faces and license plates), where the rules and requirements of the EU GDPR regulations make this especially important. However, anonymizing content unfortunately is a very slow process. An alternative to removing and/or masking actual content from the dataset for privacy reasons would be to create artificial datasets (e.g. containing public images or only faces from people who consent to publish), but this would likely also be a non-trivial task. One interesting aspect could be the use of Generative Adversarial Networks (GANs) for the anonymization of faces, for instance by replacing all faces appearing in the content with generated faces learned from a small group of people who gave their consent. Another way to preemptively mitigate the privacy issues could be to wear conspicuous ‘lifelogging stickers’ during recording to make people aware of the presence of the camera, which would give them the possibility to object to being filmed or to avoid being captured altogether.

SEPHLA: Challenges and Opportunities Within Environment-Personal Health Archives

The second presentation was given by Minh-Son Dao from the National Institute of Information and Communications Technology (NICT) in Japan about SEPHLA: Challenges and Opportunities Within Environment-Personal Health Archives [2]. This is a dataset that aims at combining the conditions of the environment with health-related aspects (e.g., pollution or weather data with cardio-respiratory or psychophysiological data). The creation of the dataset was motivated by the fact that people in larger cities in Japan very often do not want to go out (e.g., for some sports activities), because they are very concerned about pollution, i.e., health conditions. So it would be beneficial to have a map of the city with assigned pollution ratings, or a system that allows to perform related queries. Their dataset contains sensor data collected on routes by a few dozen volunteer  people over seven days in Fukuoka, Japan. More particularly, they collected data about the location, O3, NO2, PM2.5 (particulates), temperature, and humidity in combination with heart rate, motion behavior (from 3-axis accelerometer), relaxation level, and other personal perception data from questionnaires.

This dataset has also been used for multimedia benchmark challenges, such as the Lifelogging for Wellbeing task at MediaEval. In order to define the ground truth, volunteers were presented with specific use cases and annotation rules, and were asked to collaboratively annotate the dataset. The collected data (the feelings of participants at different locations) was also visualized using an interactive map. Although the dataset may have some inconsistent annotations, it is easy to filter them out since labels of corresponding annotators and annotator groups are contained in the dataset as well.

V3C – a Research Video Collection

The third presentation was given by Luca Rossetto from the University of Basel (Switzerland) about V3C – a Research Video Collection [3]. This is a large-scale dataset for multimedia retrieval, consisting of nearly 30,000 videos with an overall duration of about 3,800 hours. Although many other video datasets are available already (e.g., IACC.3 [6], or YFCC100M [8]), the V3C dataset is unique in the aspects of timeliness (more recent content than many other datasets and therefore more representative content for current ‘videos in the wild’) and diversity (represents many different genres or use cases), while also having no copyright restrictions (all contained videos were labelled with a Creative Commons license by their uploaders). The videos have been collected from the video sharing platform Vimeo (hence the name ‘Vimeo Creative Commons Collection’ or V3C in short) and represent video data currently used on video sharing platforms. The dataset comes together with a master shot-boundary detection ground truth, as well as keyframes and additional metadata. It is partitioned into three major parts (V3C1, V3C2, and V3C3) to make it more manageable, and it will be used by the TRECVID and the Video Browser Showdown (VBS) evaluation campaigns for several years. Although the dataset was not specifically built for retrieval, it is suitable for any use case that requires a larger video dataset.

The shot-boundary detection used to provide the master-shot reference for the V3C dataset was implemented using Cineast, which is an open source software available for download. It divides every frame into a 3×3 grid and computes color histograms for all 9 areas, which are then concatenated into a ‘regional color histogram’ feature vector that is compared between all adjacent frames. This seems to work very well for hard cuts and gradual transitions, although for grayscale content (and flashlights etc.) it is not very stable. The additional metadata provided with the dataset includes information about resolution, frame rate, uploading user and the upload date, as well as any semantic information provided by the uploader (title, description, tags, etc.). 

Athens Urban Soundscape (ATHUS): A Dataset for Urban Soundscape Quality Recognition

Originally a fourth presentation was scheduled about Athens Urban Soundscape (ATHUS): A Dataset for Urban Soundscape Quality Recognition [4], but unfortunately no author was on site to give the presentation. This dataset contains audio samples with a duration of 30 seconds (as well as extracted features and ground truth) from a metropolitan city (Athens, Greece), that have been recorded during a period of about four years by 10 different persons with the aim to provide a collection about city sounds. The metadata includes geospatial coordinates, timestamp, rating, and tagging of the sound by the recording person. The authors demonstrated in a baseline evaluation that their dataset allows to predict the soundscape quality in the city with about 42% accuracy.

Discussion

After the presentations, Björn Þór Jónsson moderated a panel discussion in which all presenters participated.

The panel started with a discussion on the size of datasets, whether the only way to make challenges more difficult is to keep increasing the dataset, or whether there are alternatives to this. Although this heavily depends on the research question one would like to solve, it was generally agreed that there is a definite need for evaluation with large datasets, because for small datasets some problems are trivial. Moreover, too small datasets often introduce some kind of content bias, so that they do not fully reflect the practical situation.

For now, it seems there is no real alternative to using larger datasets although it is clear that this will introduce additional challenges/hurdles for data management and data processing. All presenters (and the audience too) agreed that introducing larger datasets will also necessitate the need for closer collaboration with other research communities―with fields like data science, data management/engineering, and distributed and high-performance computing―in order to manage the higher data load.

However, even though we need larger datasets, we might not be ready yet to really go for true large-scale. For example, the V3C dataset is still far away from a true web-scale video search dataset; it originally was intended to be even bigger, but there were concerns from the TRECVID and VBS communities about the manageability. Datasets that are too large would set the entrance barrier for newcomers so high that an evaluation benchmark may not attract enough participants―a problem that could possibly disappear in a few years (as hardware becomes cheaper and faster/larger), but still needs to be addressed from an organizational viewpoint. 

There were notes from the audience that instead of focusing on size alone, we should also consider the problem we want to solve. It appears many researchers use datasets for use cases for which they were not designed and are not suited to. Instead of blindly going for larger size, datasets could be kept small and simple for solving essential research questions, for example by truly optimizing them to the problem to solve; different evaluations would then use different datasets. However, this would lead to a considerable dataset fragmentation and necessitate the need for combining several datasets for broader/larger evaluation tasks, which has been shown to be quite challenging in the past. For example, there are already a lot of health datasets available, and it would be interesting to take benefit from them, but the workload for the integration into competitions is often too high in practice.

Another issue that should be addressed more intensively by the research community is to figure out the situation for personal datasets that are compliant with GDPR regulations, since currently nobody really knows how to deal with this.

Acknowledgments

The session was organized by the authors of the report, in collaboration with Duc-Tien Dang-Nguyen (Dublin City University), Michael Riegler (Center for Digitalisation and Engineering & University of Oslo), and Luca Piras (University of Cagliari). The panel format of the special session made the discussions much more lively and interactive than that of a traditional technical session. We would like to thank the presenters and their co-authors for their excellent contributions, as well as the members of the audience who contributed greatly to the session.

References

[1] Gurrin, C., Schoeffmann, K., Joho, H., Munzer, B., Albatal, R., Hopfgartner, F., … & Dang-Nguyen, D. T. (2019, January). A test collection for interactive lifelog retrieval. In International Conference on Multimedia Modeling (pp. 312-324). Springer, Cham.
[2] Sato, T., Dao, M. S., Kuribayashi, K., & Zettsu, K. (2019, January). SEPHLA: Challenges and Opportunities Within Environment-Personal Health Archives. In International Conference on Multimedia Modeling (pp. 325-337). Springer, Cham.
[3] Rossetto, L., Schuldt, H., Awad, G., & Butt, A. A. (2019, January). V3C–A Research Video Collection. In International Conference on Multimedia Modeling (pp. 349-360). Springer, Cham.
[4] Giannakopoulos, T., Orfanidi, M., & Perantonis, S. (2019, January). Athens Urban Soundscape (ATHUS): A Dataset for Urban Soundscape Quality Recognition. In International Conference on Multimedia Modeling (pp. 338-348). Springer, Cham.
[5] Dang-Nguyen, D. T., Schoeffmann, K., & Hurst, W. (2018, June). LSE2018 Panel-Challenges of Lifelog Search and Access. In Proceedings of the 2018 ACM Workshop on The Lifelog Search Challenge (pp. 1-2). ACM.
[6] Awad, G., Butt, A., Curtis, K., Lee, Y., Fiscus, J., Godil, A., … & Kraaij, W. (2018, November). Trecvid 2018: Benchmarking video activity detection, video captioning and matching, video storytelling linking and video search.
[7] Lokoč, J., Kovalčík, G., Münzer, B., Schöffmann, K., Bailer, W., Gasser, R., … & Barthel, K. U. (2019). Interactive search or sequential browsing? a detailed analysis of the video browser showdown 2018. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 15(1), 29.
[8] Kalkowski, S., Schulze, C., Dengel, A., & Borth, D. (2015, October). Real-time analysis and visualization of the YFCC100M dataset. In Proceedings of the 2015 Workshop on Community-Organized Multimodal Mining: Opportunities for Novel Solutions(pp. 25-30). ACM.

Dataset Column: Datasets for Online Multimedia Verification

Image

By | | , ,
1 Comment

Introduction

Online disinformation is a problem that has been attracting increased interest by researchers worldwide as the breadth and magnitude of its impact is progressively manifested and documented in a number of studies (Boididou et al., 2014; Zhou & Zafarani, 2018; Zubiaga et al., 2018). This emerging area of research is inherently multidisciplinary and there have been numerous treatments of the subject, each having a distinct perspective or theme, ranging from the predominant perspectives of media, journalism and communications (Wardle & Derakhshan, 2017) and political science (Allcott & Gentzkow, 2017) to those of network science (Lazer et al., 2018), natural language processing (Rubin et al., 2015) and signal processing, including media forensics (Zampoglou et al., 2017). Given the multimodal nature of the problem, it is no surprise that the multimedia community has taken a strong interest in the field.

From a multimedia perspective, two research problems have attracted the bulk of researchers’ attention: a) detection of content tampering and content fabrication, and b) detection of content misuse for disinformation. The first was traditionally studied within the field of media forensics (Rocha et al, 2011), but has recently been under the spotlight as a result of the rise of deepfake videos (Güera & Delp, 2018), i.e. a special class of generative models that are capable of synthesizing highly convincing media content from scratch or based on some authentic seed content. The second problem has focused on the problem of multimedia misuse or misappropriation, i.e. the use of media content out of its original context with the goal of spreading misinformation or false narratives (Tandoc et al., 2018).

Developing automated approaches to detect media-based disinformation is relying to a great extent on the availability of relevant datasets, both for training supervised learning models and for evaluating their effectiveness. Yet, developing and releasing such datasets is a challenge in itself for a number of reasons:

  1. Identifying, curating, understanding, and annotating cases of media-based misinformation is a very effort-intensive task. More often than not, the annotation process requires careful and extensive reading of pertinent news coverage from a variety of sources similar to the journalistic practice of verification (Brandtzaeg et al., 2016).
  2. Media-based disinformation is largely manifested in social media platforms and relevant datasets are therefore hard to collect and distribute due to the temporary nature of social media content and the numerous technical restrictions and challenges involved in collecting content (mostly due to limitations or complete lack of appropriate support by the respective APIs), as well as the legal and ethical issues in releasing social media-based datasets (due to the need to comply with the respective Terms of Service and any applicable data protection law).

In this column, we present two multimedia datasets that could be of value to researchers who study media-based disinformation and develop automated approaches to tackle the problem. The first, called Fake Video Corpus (Papadopoulou et al., 2019) is a manually curated collection of 200 debunked and 180 verified videos, along with relevant annotations, accompanied by a set of 5,193 near-duplicate instances of them that were posted on popular social media platforms. The second, called FIVR-200K (Kordopatis-Zilos et al., 2019), is an automatically collected dataset of 225,960 videos, a list of 100 video queries and manually verified annotations regarding the relation (if any) of the dataset videos to each of the queries (i.e. near-duplicate, complementary scene, same incident).

For each of the two datasets, we present the design and creation process, focusing on issues and questions regarding the relevance of the collected content, the technical means of collection, and the process of annotation, which had the dual goal of ensuring high accuracy and keeping the manual annotation cost manageable. Given that each dataset is accompanied by a detailed journal article, in this column we only limit our description to high-level information, emphasizing the utility and creation process in each case, rather than on detailed statistics, which are disclosed in the respective papers.

Following the presentation of the two datasets, we then proceed to a critical discussion, highlighting their limitations and some caveats, and delineating future steps towards high quality dataset creation for the field of multimedia-based misinformation.

Related Datasets

The complexity and challenge of the multimedia verification problem has led to the creation of numerous datasets and benchmarking efforts, each designed specifically for a particular task within this area. We can broadly classify these efforts in three areas: a) multimedia forensics, b) multimedia retrieval, and c) multimedia post classification. Datasets that are focused on the text modality, e.g. Fake News Challenge, Clickbait Challenge, Hyperpartisan News Detection, RumourEval (Derczynski et al 2017), etc. are beyond the scope of this post and are hence not included in this discussion.

Multimedia forensics: Generating high-quality multimedia forensics datasets has always been a challenge, since creating convincing forgeries is normally a manual task requiring a fair amount of skill, and as a result such datasets have generally been few and limited in scale. With respect to image splicing, our own survey (Zampoglou et al, 2017) listed a number of datasets that had been made available by this point, including our own Wild Web tampered image dataset, which consists of real-world forgeries that have been collected from the Web, including multiple near-duplicates, making it a large and particularly challenging collection. Recently, the Realistic Tampering Dataset (Korus et al,2017) was proposed, offering a large number of convincing forgeries for evaluation. On the other hand, copy-move image forgeries pose a different problem that requires specially designed datasets. Three such commonly used datasets are those produced by MICC (Amerini et al, 2011), the Image Manipulation Dataset by (Christlein et al, 2012), and CoMoFoD (Tralic et al, 2013). These datasets are still actively used in research.

With respect to video tampering, there has been relative scarcity in high-quality large-scale datasets, which is understandable given the difficulty of creating convincing forgeries. The recently proposed Multimedia Forensics Challenge datasets include some large-scale sets of tampered images and videos for the evaluation of forensics algorithms. Finally, there has recently been increased interest towards the automatic detection of forgeries made with the assistance of particular software, and specifically face-swapping software. As the quality of produced face-swaps is constantly improving, detecting face-swaps is an important emerging verification task. The FaceForensics++ dataset (Rössler et al, 2019) is a very-large scale dataset containing face-swapped videos (and untampered face videos) from a number of different algorithms, aimed for the evaluation of face-swap detection algorithms.

Multimedia retrieval: Several cases of multimedia verification can be considered to be an instance of a near-duplicate retrieval task, in which the query video (video to be verified) is run against a database of past cases/videos to check whether it has already appeared before. The most popular and publicly-available dataset for near-duplicate video retrieval is arguably the CC_WEB_VIDEO dataset (Wu et al., 2007). This consists of 12,790 user-generated videos collected from popular video sharing websites (YouTube, Google Video, and Yahoo! Video). It is organized in 24 query sets, for each of which the most popular video was selected to serve as query, and the rest of the videos were manually annotated based on their duplicity to the query. Another relevant dataset is VCDB (Jiang et al., 2014), which was compiled and annotated as a benchmark for the partial video copy detection problem and is composed of videos from popular video platforms (YouTube and Metacafe). VCDB contains two subsets of videos: a) the core, which consists of 28 discrete sets of videos with a total of 528 videos with over 9,000 pairs of manually annotated partial copies, and b) the distractors, which consists of 100,000 videos with the purpose to make the video copy detection problem more challenging.

Multimedia post classification: A benchmark task under the name “Verifying Multimedia Use” (Boididou et al., 2015; Boididou et al., 2016) was organized and took place in the context of MediaEval 2015 and 2016 respectively. The task made a dataset available of 15,629 tweets containing images and videos, each of which made a false or factual claim with respect to the shared image/video. The released tweets were posted in the context of breaking news events (e.g. Hurricane Sandy, Boston Marathon bombings) or hoaxes. 

Video Verification Datasets

The Fake Video Corpus (FVC)

The Fake Video Corpus (Papadopoulou et al., 2018) is a collection of 380 user-generated videos and 5,193 near-duplicate versions of them, all collected from three online video platforms: YouTube, Facebook, and Twitter. The videos are annotated either as “verified” (“real”) or as “debunked” (“fake”) depending on whether the information they convey is accurate or misleading. Verified videos are typically user-generated takes of newsworthy events, while debunked videos include various types of misinformation, including staged content posing as UGC, real content taken out of context, or modified/tampered content (see Figure 1 for examples). The near-duplicates of each video are arranged in temporally ordered “cascades”, and each near-duplicate video is annotated with respect to its relation to the first video of the cascade (e.g. whether it is reinforcing or debunking the original claim). The FVC is the first, to our knowledge, large-scale dataset of debunked and verified user-generated videos (UGVs). The dataset contains different kinds of metadata for its videos, including channel (user) information, video information, and community reactions (number of likes, shares and comments) at the time of their inclusion.

  
  
Figure 1. A selection of real (top row) and fake (bottom row) videos from the Fake Video Corpus. Click image to jump to larger version, description, and link to YouTube video.

The initial set of 380 videos were collected and annotated using various sources including the Context Aggregation and Analysis (CAA) service developed within the InVID project and fact-checking sites such as Snopes. To build the dataset, all videos submitted to the CAA service between November 2017 and January 2018 were collected in an initial pool of approximately 1600 videos, which were then manually inspected and filtered. The remaining videos were annotated as “verified” or “debunked” using established third party sources (news articles or blog posts), leading to the final pool of 180 verified and 200 fake unique videos. Then, keyword-based search was run on the three platforms, and near-duplicate video detection was used to identify the video duplicates within the returned results. More specifically, for each of the 380 videos, its title was reformulated in a more general form, and translated into four major languages: Russian, Arabic, French, and German. The original title, the general form and the translations were submitted as queries to YouTube, Facebook, and Twitter. Then, the  near-duplicate retrieval algorithm of Kordopatis-Zilos etal (2017) was used on the resulting pool, and the results were manually inspected to remove erroneous matches.

The purpose of the dataset is twofold: i) to be used for the analysis of the dissemination patterns of real and fake user-generated videos (by analyzing the traits of the near-duplicate video cascades), and ii) to serve as a benchmark for the evaluation of automated video verification methods. The relatively large size of the dataset is important for both of these tasks. With respect to the study of dissemination patterns, the dataset provides the opportunity to study the dissemination of the same or similar content by analyzing associations between videos not provided by the original platform APIs, combined with the wealth of associated metadata. In parallel, having a collection of 5,573 annotated “verified” or “debunked” videos- even if many are near-duplicate versions of the 380 cases – can be used for the evaluation (or even training) of verification systems, either based on visual content or the associated video metadata.

The Fine-grained Incident Video Retrieval Dataset (FIVR-200K)

The FIVR-200K dataset (Kordopatis-Zilos et al., 2019) consists of 225,960 videos associated with 4,687 Wikipedia events and 100 selected video queries (see Figure 2 for examples). It has been designed to simulate the problem of Fine-grained Incident Video Retrieval (FIVR). The objective of this problem is: given a query video, retrieve all associated videos considering several types of associations with respect to an incident of interest. FIVR contains several retrieval tasks as special cases under a single framework. In particular, we consider three types of association between videos: a) Duplicate Scene Videos (DSV), which share at least one scene (originating from the same camera) regardless of any applied transformation, b) Complementary Scene Videos (CSV), which contain part of the same spatiotemporal segment, but captured from different viewpoints, and c) Incident Scene Videos (ISV), which capture the same incident, i.e. they are spatially and temporally close, but have no overlap.

For the collection of the dataset, we first crawled Wikipedia’s Current Event page to collect a large number of major news events that occurred between 2013 and 2017 (five years). Each news event is accompanied with a topic, headline, text, date, and hyperlinks. To collect videos of the same category, we retained only news events with topic “Armed conflicts and attacks” or “Disasters and accidents”. This ultimately led to a total of 4,687 events after filtering. To gather videos around these events and build a large collection with numerous video pairs that are associated through the relations of interest (DSV, CSV and ISV), we queried the public YouTube API with the event headlines. To ensure that the collected videos capture the corresponding event, we retained only the videos published within a timespan of one week from the event date. This process resulted in the collection of 225,960 videos.

  
Figure 2. A selection of query videos from the Fine-grained Incident Video Retrieval dataset. Click image to jump to larger version, link to YouTube video, and several associated videos.

Next, we proceeded with the selection of query videos. We set up an automated filtering and ranking process that implemented the following criteria: a) query videos should be relatively short and ideally focus on a single scene, b) queries should have many near-duplicates or same-incident videos within the dataset that are published by many different uploaders, c) among a set of near-duplicate/same-instance videos, the one that was uploaded first should be selected as query. This selection process was implemented based on a graph-based clustering approach and resulted in the selection of 635 query videos, of which we used the top 100 (ranked by corresponding cluster size) as the final query set.

For the annotation of similarity relations among videos, we followed a multi-step process, in which we presented annotators with the results of a similarity-based video retrieval system and asked them to indicate the type of relation through a drop-down list of the following labels: a) Near-Duplicate (ND), a special case where the whole video is near-duplicate to the query video, b) Duplicate Scene (DS), where only some scenes in the candidate video are near-duplicates of scenes in the query video, c) Complementary Scenes (CS), d) Incident Scene (IS), and e) Distractors (DI), i.e. irrelevant videos.

To make sure that annotators were presented with as many potentially relevant videos as possible, we used visual-only, text-only and hybrid similarity in turn. As a result, each annotator reviewed video candidates that had very high similarity with the query video in terms either of their visual content, or text metadata (title and description) or the combination of similarities. Once an initial set of annotations were produced by two independent annotators, the annotators went twice again through the annotations two ensure consistency and accuracy.

FIVR-200K was designed to serve as a benchmark that poses real-world challenges for the problem of reverse video search. Given a query video to be verified, the analyst would want to know whether the same or a very similar version of it has already been published. In that way, the user would be able to easily debunk cases of out-of-context video use (i.e. misappropriation) and on the other hand, if several videos are found that depict the same scene from different viewpoints at approximately the same time, then they could be considered to corroborate the video of interest.

Discussion: Limitations and Caveats

We are confident that the two video verification datasets presented in this column can be valuable resources for researchers interested in the problem of media-based disinformation and could serve both as training sets and as benchmarks for automated video verification methods. Yet, both of them suffer from certain limitations and care should be taken when using them to draw conclusions. 

A first potential issue has to do with the video selection bias arising from the particular way that each of the two datasets was created. The videos of the Fake Video Corpus were selected in a mixed manner trying to include a number of cases that were known to the dataset creators and their collaborators, and was also enriched by a pool of test videos that were submitted for analysis to a publicly available video verification service. As a result, it is likely to be more focused on viral and popular videos. Also, videos were included, for which debunking or corroborating information was found online, which introduces yet another source of bias, potentially towards cases that were more newsworthy or clear cut. In the case of the FIVR-200K dataset, videos were intentionally collected to be between two categories of newsworthy events with the goal of ending up with a relatively homogeneous collection, which would be challenging in terms of content-based retrieval. This means that certain types of content, such as political events, sports and entertainment, are very limited or not present at all in the dataset. 

A question that is related to the selection bias of the above datasets pertains to their relevance for multimedia verification and for real-world applications. In particular, it is not clear whether the video cases offered by the Fake Video Corpus are representative of actual verification tasks that journalists and news editors face in their daily work. Another important question is whether these datasets offer a realistic challenge to automatic multimedia analysis approaches. In the case of FIVR-200K, it was clearly demonstrated (Kordopatis-Zilos et al., 2019) that the dataset is a much harder benchmark for near-duplicate detection methods compared to previous datasets such as CC_WEB_VIDEO and VCDB. Even so, we cannot safely conclude that a method, which performs very well in FIVR-200K, would perform equally well in a dataset of much larger scale (e.g. millions or even billions of videos).

Another issue that affects the access to these datasets and the reproducibility of experimental results relates to the ephemeral nature of online video content. A considerable (and increasing) part of these video collections is taken down (either by their own creators or from the video platform), which makes it impossible for researchers to gain access to the exact video set that was originally collected. To give a better sense of the problem, 21% of the Fake Video Corpus and 11% of the FIVR-200K videos were not available online on September 2019. This issue, which affects all datasets that are based on online multimedia content, raises the more general question of whether there are steps that can be taken by online platforms such as YouTube, Facebook and Twitter that could facilitate the reproducibility of social media research without violating copyright legislation or the platforms’ terms of service.

The ephemeral nature of online content is not the only factor that renders the value of multimedia datasets very sensitive to the passing of time. Especially in the case of online disinformation, there seems to be an arms’ race, where new machine learning methods constantly get better in detecting misleading or tampered content, but at the same time new types of misinformation emerge, which are increasingly AI-assisted. This is particularly profound in the case of deepfakes, where the main research paradigm is based on the concept of competition between a generator (adversary) and a detector (Goodfellow et al., 2014). 

Last but not least, one may always be concerned about the potential ethical issues arising when publicly releasing such datasets. In our case, reasonable concerns for privacy risks, which are always relevant when dealing with social media content, are addressed by complying with the relevant Terms of Service of the source platforms and by making sure that any annotation (label) assigned to the dataset videos is accurate. Additional ethical issues pertain to the potential “dual use” of the dataset, i.e. their use by adversaries to craft better tools and techniques to make misinformation campaigns more effective. A recent pertinent case was OpenAI’s delayed release of their very powerful GPT-2 model, which sparked numerous discussions and criticism, and making clear that there is no commonly accepted practice for ensuring reproducibility of research results (and empowering future research) and at the same time making sure that risks of misuse are eliminated.

Future work

Given the challenges of creating and releasing a large-scale dataset for multimedia verification, the main conclusions from our efforts towards this direction so far are the following:

  • The field of multimedia verification is in constant motion and therefore the concept of a static dataset may not be sufficient to capture the real-world nuances and latest challenges of the problem. Instead new benchmarking models, e.g. in the form of open data challenges, and resources, e.g. constantly updated repository of “fake” multimedia, appear to be more effective for empowering future research in the area.
  • The role of social media and multimedia sharing platforms (incl. YouTube, Facebook, Twitter, etc.) seems to be crucial in enabling effective collaboration between academia and industry towards addressing the real-world consequences of online misinformation. While there have been recent developments towards this direction, including the announcements by both Facebook and Alphabet’s Jigsaw of new deepfake datasets, there is also doubt and scepticism about the degree of openness and transparency that such platforms are ready to offer, given the conflicts of interest that are inherent in the underlying business model. 
  • Building a dataset that is fit for a highly diverse and representative set of verification cases appears to be a task that would require a community effort instead of effort from a single organisation or group. This would not only help towards distributing the massive dataset creation cost and effort to multiple stakeholders, but also towards ensuring less selection bias, richer and more accurate annotation and more solid governance.

References

Allcott, H., Gentzkow, M., “Social media and fake news in the 2016 election”, Journal of economic perspectives, 31(2), pp. 211–36, 2017.
Amerini, I, Ballan, L., Caldelli, R., Del Bimbo, A., Serra, G., “A SIFT-based forensic method for copy-move attack detection and transformation recovery”, IEEE Transactions on Information Forensics and Security, 6(3), pp. 1099–1110,2011.
Boididou, C., Papadopoulos, S., Kompatsiaris, Y., Schifferes, S., Newman, N., “Challenges of computational verification in social multimedia”, In Proceedings of the 23rd ACM International Conference on World Wide Web, pp. 743–748,2014.
Boididou, C., Andreadou, K., Papadopoulos, S., Dang-Nguyen, D.T., Boato, G., Riegler, M., Kompatsiaris, Y., “Verifying multimedia use at MediaEval 2015”. In Proceedings of MediaEval 2015, 2015.
Boididou C., Papadopoulos S., Dang-Nguyen D., Boato G., Riegler M., Middleton S.E., Petlund A., Kompatsiaris Y., “Verifying multimedia use at MediaEval 2016”. In Proceedings of MediaEval 2016, 2016.
Brandtzaeg, P.B., Lüders, M., Spangenberg, J., Rath-Wiggins, L., Følstad, A., “Emerging journalistic verification practices concerning social media”. Journalism Practice, 10(3), pp. 323–342, 2016.
Christlein V., Riess C., Jordan J., Riess C., Angelopoulou, E., “An evaluation of popular copy-move forgery detection approaches”. IEEE Transactions on Information Forensics & Security, 7(6), pp. 1841–1854, 2012.
Derczynski, L., Bontcheva, K., Liakata, M., Procter, R., Hoi, G.W.S., Zubiaga, A., “Semeval-2017 Task 8: Rumoureval: determining rumour veracity and support for rumours”, Proceedings of the 11th International Workshop on Semantic Evaluation,pp. 69-76, 2017.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Bengio, Y., “Generative adversarial nets”. In Advances in Neural Information Processing Systems, pp. 2672–2680, 2014.
Guan, H., Kozak, M., Robertson, E., Lee, Y., Yates, A.N., Delgado, A., Zhou, D., Kheyrkhah, T., Smith, J., Fiscus, J., “MFC datasets: Large-scale benchmark datasets for media forensic challenge evaluation”, In Proceedings of the 2019 IEEEWinter Applications of Computer Vision Workshops, pp. 63–72, 2019.
Güera, D., Delp, E.J., “Deepfake video detection using recurrent neural networks”, In Proceedings of the 15th IEEE International Conference on Advanced Video and Signal Based Surveillance, pp. 1–6, 2018.
Jiang, Y. G., Jiang, Y., Wang, J., “VCDB: A large-scale database for partial copy detection in videos”. In Proceedings of the European Conference on Computer Vision, pp. 357–371, 2014.
Kiesel, J., Mestre, M., Shukla, R., Vincent, E., Adineh, P., Corney, D., Stein, B. Potthast, M., “Semeval-2019 Task 4: Hyperpartisan news detection”. In Proceedings of the 13th International Workshop on Semantic Evaluation, pp. 829–839,2019.
Kordopatis-Zilos, G., Papadopoulos, S., Patras, I., Kompatsiaris, I., “FIVR: Fine-grained incident video retrieval”. IEEE Transactions on Multimedia, 21(10), pp. 2638–2652, 2019.
Korus, P., Huang, J., “Multi-scale analysis strategies in PRNU-based tampering localization”, IEEE Transactions on Information Forensics & Security, 21(4), pp. 809–824, 2017.
Lazer, D.M., Baum, M.A., Benkler, Y., Berinsky, A.J., Greenhill, K.M., Menczer, F., Schudson, M., “The science of fake news”, Science, 359(6380), pp. 1094–1096, 2018.
Papadopoulou, O., Zampoglou, M., Papadopoulos, S., Kompatsiaris, I., “A corpus of debunked and verified user-generated videos”. Online Information Review, 43(1), pp. 72–88, 2019.
Rocha, A., Scheirer, W., Boult, T., Goldenstein, S., “Vision of the unseen: Current trends and challenges in digital image and video forensics”, ACM Computing Surveys, 43(4), art. 26, 2011.
Rössler, A. Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Nießner, M. “Faceforensics++: Learning to detect manipulated facial images”, In Proceedings of the IEEE International Conference on Computer Vision, 2019.
Rubin, V.L., Chen, Y., Conroy, N.J., “Deception detection for news: Three types of fakes”, In Proceedings of the 78th ASIS&T Annual Meeting: Information Science with Impact: Research in and for the Community, art. 83, 2015.
Tandoc Jr, E.C., Lim, Z.W., Ling, R. “Defining “fake news”: A typology of scholarly definitions”, Digital journalism, 6(2), pp. 137–153, 2018.
Tralic, D., Zupancic I., Grgic S., Grgic M., “CoMoFoD – New database for copy-move forgery detection”. In Proceedings of the 55th International Symposium on Electronics in Marine, pp. 49–54, 2013.
Wardle, C., Derakhshan, H., “Information disorder: Toward an interdisciplinary framework for research and policy making”, Council of Europe Report, 27, 2017.
Wu, X., Hauptmann, A.G., Ngo, C.-W., “Practical elimination of near-duplicates from web video search”, In Proceedings of the 15th ACM International Conference on Multimedia, pp. 218–227, 2007.
Zampoglou, M., Papadopoulos, S., Kompatsiaris, Y., “Detecting image splicing in the wild (web)”, In Proceedings of the 2015 IEEE International Conference on Multimedia & Expo Workshops, 2015.
Zampoglou, M., Papadopoulos, S., Kompatsiaris, Y., “Large-scale evaluation of splicing localization algorithms for web images”, Multimedia Tools and Applications, 76(4), pp. 4801–4834, 2017.
Zhou, X., Zafarani, R., “Fake news: A survey of research, detection methods, and opportunities”. arXiv preprint arXiv:1812.00315, 2018.
Zubiaga, A., Aker, A., Bontcheva, K., Liakata, M., Procter, R., “Detection and resolution of rumours in social media: A survey”, ACM Computing Surveys, 51(2), art. 32, 2018.

Appendix A: Examples of videos in the Fake Video Corpus.

Real videos


US Airways Flight 1549 ditched in the Hudson River.


A group of musicians playing in an Istanbul park while bombs explode outside the stadium behind them.


A giant alligator crossing a Florida golf course.

Fake videos


“Syrian boy rescuing a girl amid gunfire” – Staged (fabricated content): The video was filmed by Norwegian Lars Klevberg in Malta.


“Golden Eagle Snatches Kid” – Tampered: The video was created by a team of students in Montreal as part of their course on visual effects.


“Pope Francis slaps Donald Trump’s hand for touching him” – Satire/parody: The video was digitally manipulated, and was made for the late-night television show Jimmy Kimmel Live.

Appendix B: Examples of videos in the Fine-grained Incident Video Retrieval dataset.

Example 1


Query video from the American Airlines Flight 383 fire at Chicago O’Hare International Airport in October 28, 2016.


Duplicate scene video.


Complimentary scene video.


Incident scene video.

Example 2


Query video from the Boston Marathon bombing in April 15, 2013.


Duplicate scene video.


Complimentary scene video.


Incident scene video.

Example 3


Query video from the the Las Vegas shooting in October 1, 2017.


Duplicate scene video.


Complimentary scene video.


Incident scene video.

Datasets and Benchmarks Column: Introduction

By | | ,
Leave a comment

Datasets are critical for research and development as, rather obviously, data is required for performing experiments, validating hypotheses, analyzing designs, and building applications. Over the years a plurality of multimedia collections have been put together, which can range from the one-off instances that have been exclusively created for supporting the work presented in a single paper or demo to those that have been created with multiple related or separate endeavors in mind. Unfortunately, the collected data is often not made publicly available. In some cases, it may not be possible to make a public release due to the proprietary or sensitive nature of the data, but other forces are also at work. For example, one might be reluctant to share data freely, as it has a value from the often substantial amount of time, effort, and money that was invested in collecting it. 

Once a dataset has been made public though, it becomes possible to perform validations of results reported in the literature and to make comparisons between methods using the same source of truth, although matters are complicated when the source code of the methods is not published or the ground truth labels are not made available. Benchmarks offer a useful compromise by offering a particular task to solve along with the data that one is allowed to use and the evaluation metrics that dictate what is considered success and failure. While benchmarks may not offer the cutting edge of research challenges for which utilizing the freshest data is an absolute requirement, they are a useful sanity check to ensure that methods that appear to work on paper also work in practice and are indeed as good as claimed.

Several efforts are underway to stimulate sharing of datasets and code, as well as to promote the reproducibility of experiments. These efforts provide encouragement to overcome the reluctance to share data by underlining the ways in which data becomes more valuable with community-wide use. They also offer insights on how researchers can put data sets that are publicly available to the best possible use. We provide here a couple of key examples of ongoing efforts. At the MMSys conference series, there is a special track for papers on datasets, and Qualinet maintains an index of known multimedia collections. The ACM Artifact Review and Badging policy proposal recommends journals and conferences to adopt a reviewing procedure where the submitted papers can be granted special badges to indicate to what extent the performed experiments are repeatable, replicable, and reproducible. For example, the “Artifacts Evaluated – Reusable” badge would indicate that artifacts associated with the research are found to be documented, consistent, complete, exercisable, and include appropriate evidence of verification and validation to the extent that reuse and repurposing is facilitated.

In future posts appearing in this column, we will be highlighting new public datasets and upcoming benchmarks through a series of invited guest posts, as well as provide insights and updates on the latest development in this area. The columns are edited by Bart Thomee and Martha Larson (see our bios at the end of this post).

To establish a baseline of popular multimedia datasets and benchmarks that have been used over the years by the research community, refer to the table below to see what the state of the art was as of 2015 when the data was compiled by Bart for his paper on the YFCC100M dataset. We can see the sizes of the datasets steadily increasing over the years, the license becoming less restrictive, and it now is the norm to also release additional metadata, precomputed features, and/or ground truth annotations together with the dataset. The last three entries in the table are benchmarks that include tasks such as video surveillance and object localization (TRECVID), diverse image search and music genre recognition (MediaEval), life-logging event search and medical image analysis (ImageCLEF), to name just a few. The table is most certainly not exhaustive, although it is reflective of the evolution of datasets over the last two decades. We will use this table to provide context for the datasets and benchmarks that we will cover in our upcoming columns, so stay tuned for our next post!

collections

bartBart Thomee is a Software Engineer at Google/YouTube in San Bruno, CA, USA, where he focuses on web-scale real-time streaming and batch techniques to fight abuse, spam, and fraud. He was previously a Senior Research Scientist at Yahoo Labs and Flickr, where his research centered on the visual and spatiotemporal dimensions of media, in order to better understand how people experience and explore the world, and how to better assist them with doing so. He led the development of the YFCC100M dataset released in 2014, and previously was part of the efforts leading to the creation of both MIRFLICKR datasets. He has furthermore been part of the organization of the ImageCLEF photo annotation tasks 2012–2013, the MediaEval placing tasks 2013–2016, and the ACM MM Yahoo-Flickr Grand Challenges 2015–2016. In addition, he has served on the program committees of, amongst others, ACM MM, ICMR, SIGIR, ICWSM and ECIR. He was part of the Steering Committee of the Multimedia COMMONS 2015 workshop at ACM MM and co-chaired the workshop in 2016; he also co-organized the TAIA workshop at SIGIR 2015.

Martha Larson is professor in the area of multiSquaremedia information technology at Radboud University in Nijmegen, Netherlands. Previously, she researched and lectured in the area of audio-visual retrieval Fraunhofer IAIS, Germany, and at the University of Amsterdam, Netherlands. Larson is co-founder of the MediaEval international benchmarking initiative for Multimedia Evaluation. She has contributed to the organization of various other challenges, including CLEF NewsREEL 2015-2017, ACM RecSys Challenge 2016, and TRECVid Video Hyperlinking 2016. She has served on the program committees of numerous conferences in the area of information retrieval, multimedia, recommender systems, and speech technology. Other forms of service have included: Area Chair at ACM Multimedia 2013, 2014, 2017, and TPC Chair at ACM ICMR 2017. Currently, she is an Associated Editor for IEEE Transactions of Multimedia. She is a founding member of the ISCA Special Interest Group on Speech and Language in Multimedia and serves on the IAPR Technical Committee 12 Multimedia and Visual Information Systems. Together with Hayley Hung she developed and currently teaches an undergraduate course in Multimedia Analysis at Delft University of Technology, where she maintains a part-time membership in the Multimedia Computing Group.