Thierry Malon is a PhD student in informatics and computer vision at the Institut de Recherche en Informatique de Toulouse (IRIT). His research focuses on pattern recognition, scene analysis, and interactive video search.
Geoffrey Roman Jimenez is a Post-Doctoral Fellow in data science at the Institut de Recherche en Informatique de Toulouse (IRIT). His research concerns statistical modeling, data representation and machine-/transfer-learning.
Patrice Guyot is an Associate Professor at IMT Mines Alès. His research focuses on audio signals and human perception.
Sylvie Chambon is an Associate Professor at the Institut de Recherche en Informatique de Toulouse (IRIT).
Vincent Charvillat is a Full Professor in computer science at the Institut de Recherche en Informatique de Toulouse (IRIT).
Alain Crouzil is an Associate Professor at the Institut de Recherche en Informatique de Toulouse (IRIT). His research interests include computer vision, image and video analysis.
André Péninou is an Associate Professor at the Institut de Recherche en Informatique de Toulouse (IRIT). His research concerns heterogeneous data modeling and management, and data access.
Julien Pinquier is an Associate Professor at the Institut de Recherche en Informatique de Toulouse (IRIT). His research focuses on audio segmentation, machine learning and automatic structuring of audiovisual documents.
Florence Sèdes is a Full Professor in Computer Science. Her research centers on data science and AI, with applications dedicated to metadata, social media analysis, security and CCTV.
Christine Sénac is an Associate Professor at the Institut de Recherche en Informatique de Toulouse (IRIT). Her research focuses on audio and audio/video content analysis and structuring.
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 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” . 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)  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)  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)  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  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 ||VIRAT ||MuHAVi ||Human3.6M ||ToCaDa |
|# Cameras||8 static||16 static||8 static||4 static||25 static|
|Overlapping FOV||Very partially||2+2||8||4||17|
|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|
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  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.
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.
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.
A special annotation procedure was set to handle the audiovisual content of this multi-view data . 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.
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).
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.
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.
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.
 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.
 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.
 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.
 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.
 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.
 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.
 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.