State-of-the-art temporal action detectors inefficiently search the entire video for specific actions. Despite the encouraging progress these methods achieve, it is crucial to design automated approaches that only explore parts of the video which are the most relevant to the actions being searched. To address this need, we propose the new problem of action spotting in videos, which we define as finding a specific action in a video while observing a small portion of that video. Inspired by the observation that humans are extremely efficient and accurate in spotting and finding action instances in a video, we propose Action Search, a novel Recurrent Neural Network approach that mimics the way humans spot actions. Moreover, to address the absence of data recording the behavior of human annotators, we put forward the Human Searches dataset, which compiles the search sequences employed by human annotators spotting actions in the AVA and THUMOS14 datasets. We consider temporal action localization as an application of the action spotting problem. Experiments on the THUMOS14 dataset reveal that our model is not only able to explore the video efficiently (observing on average 17.3% of the video) but it also accurately finds human activities with 30.8% mAP (0.5 tIoU), outperforming state-of-the-art methods

We are stumbling across a video tsunami flooding our communication channels. The ubiquity of digital cameras and social networks has increased the amount of visual media content generated and shared by people, in particular videos. Cisco reports that 82% of the internet traffic would be in the form of videos by 2022. The computer vision community has embraced this challenge by offering the first building blocks to translate the visual data in segmented video clips into semantic tags. However, users usually require to go beyond tagging at the video level. For example, someone may want to retrieve important moments such as the “first steps of her child” from a large collection of untrimmed videos; or retrieving all the instances of a home-run from an unsegmented video of baseball. In the face of this data deluge, it becomes crucial to develop efficient and scalable algorithms that can intelligently localize semantic visual content in untrimmed videos. In this work, I address three different challenges on the localization of actions in videos. First, I develop deep-based action proposals and detection models that take a video and generate action-agnostic and class-specific temporal segments, respectively. These models retrieve temporal locations with high accuracy in an efficient manner, faster than real-time. Second, I propose the new task to retrieve and localize temporal moments from a collection of videos given a natural language query. To tackle this challenge, I introduce an efficient and effective model that aligns the text query to individual clips of fixed length while still retrieves moments spanning multiple clips. This approach not only allows smooth interactions with users via natural languagequeries but also reduce the index size and search time for retrieving the moments. Lastly, I introduce the concept of actor-supervision that exploits the inherent compo sitionality of actions, in terms of transformations of actors, to achieve spatiotemporal localization of actions without the need of action box annotations. By designing ef ficient models to scan a single video in real-time; retrieve and localizing moments of interest from multiple videos; and an effective strategy to localize actions without resorting in action box annotations, this thesis provides insights that put us closer to the goal of general video understanding.

In this thesis the problem of automatic human action recognition and localization in videos is studied. In this problem, our goal is to recognize the category of the human action that is happening in the video, and also to localize the action in space and/or time. This problem is challenging due to the complexity of the human actions, the large intra-class variations and the distraction of backgrounds. Human actions are inherently structured patterns of body movements. However, past works are inadequate in learning the space-time structures in human actions and exploring them for better recognition and localization. In this thesis new methods are proposed that exploit such space-time structures for effective human action recognition and localization in videos, including sports videos, YouTube videos, TV programs and movies. A new local space-time video representation, the hierarchical Space-Time Segments, is first proposed. Using this new video representation, ensembles of hierarchical spatio-temporal trees, discovered directly from the training videos, are constructed to model the hierarchical, spatial and temporal structures of human actions. This proposed approach achieves promising performances in action recognition and localization on challenging benchmark datasets. Moreover, the discovered trees show good cross-dataset generalizability: trees learned on one dataset can be used to recognize and localize similar actions in another dataset. To handle large scale data, a deep model is explored that learns temporal progression of the actions using Long Short Term Memory (LSTM), which is a type of Recurrent Neural Network (RNN). Two novel ranking losses are proposed to train the model to better capture the temporal structures of actions for accurate action recognition and temporal localization. This model achieves state-of-art performance on a large scale video dataset. A deep model usually employs a Convolutional Neural Network (CNN) to learn visual features from video frames. The problem of utilizing web action images for training a Convolutional Neural Network (CNN) is also studied: training CNN typically requires a large number of training videos, but the findings of this study show that web action images can be utilized as additional training data to significantly reduce the burden of video training data collection.

The rise of deep learning has facilitated remarkable progress in video understanding. This thesis addresses three important tasks of video understanding: video object detection, joint object and action detection, and spatio-temporal action localization. Object class detection is one of the most important challenges in computer vision. Object detectors are usually trained on bounding-boxes from still images. Recently, video has been used as an alternative source of data. Yet, training an object detector on one domain (either still images or videos) and testing on the other one results in a significant performance gap compared to training and testing on the same domain. In the first part of this thesis, we examine the reasons behind this performance gap. We define and evaluate several domain shift factors: spatial location accuracy, appearance diversity, image quality, aspect distribution, and object size and camera framing. We examine the impact of these factors by comparing the detection performance before and after cancelling them out. The results show that all five factors affect the performance of the detectors and their combined effect explains the performance gap. While most existing approaches for detection in videos focus on objects or human actions separately, in the second part of this thesis we aim at detecting non-human centric actions, i.e., objects performing actions, such as cat eating or dog jumping. We introduce an end-to-end multitask objective that jointly learns object-action relationships. We compare it with different training objectives, validate its effectiveness for detecting object-action pairs in videos, and show that both tasks of object and action detection benefit from this joint learning. In experiments on the A2D dataset [Xu et al., 2015], we obtain state-of-the-art results on segmentation of object-action pairs. In the third part, we are the first to propose an action tubelet detector that leverages the temporal continuity of videos instead of operating at the frame level, as state-of-the-art approaches do. The same way modern detectors rely on anchor boxes, our tubelet detector is based on anchor cuboids by taking as input a sequence of frames and outputing tubelets, i.e., sequences of bounding boxes with associated scores. Our tubelet detector outperforms all state of the art on the UCF-Sports [Rodriguez et al., 2008], J-HMDB [Jhuang et al., 2013a], and UCF-101 [Soomro et al., 2012] action localization datasets especially at high overlap thresholds. The improvement in detection performance is explained by both more accurate scores and more precise localization.

Video action understanding involves comprehending actions performed by humans, depicted in videos. Central to the task of video action understanding are four fundamental questions: What, When, Where, and Who. These questions encapsulate the essence of action classification, temporal action localization, action detection, and actor recognition. Despite notable progress in research related to these tasks, many challenges persist and in this dissertation, we propose innovative solutions to tackle these challenges head-on. First, we address the challenges in action classification (``What?"), specifically related to multi-view action recognition. We propose a novel transformer decoder-based model, with learnable view and action queries, to enforce the learning of action features robust to shifts in viewpoints. Next, we focus on temporal action localization (``What?" and ``When?") and address challenges introduced in the multi-label setting. Our proposed solution involves leveraging the inherent relationships between complex actions in real-world videos. We introduce an attention-based architecture that models these relationships for the task of temporal action localization. Next, we propose \textit{Gabriella}, a real-time online system for activity detection (``What?", ``When?", and ``Where?") in security videos. Our proposed solution has three stages: tubelet extraction, activity classification, and online tubelet merging. For tubelet extraction, we propose a localization network that detects potential foreground regions to generate action tubelets. The detected tubelets are assigned activity class scores by the classification network and merged using our proposed Tubelet-Merge Action-Split (TMAS) algorithm to form the final action detections. Finally, we introduce an approach to solve the novel task of joint action and actor recognition (``What?" and ``Who?") and solve it using disentangled representation learning. We introduce a novel method to simultaneously identify both subjects (actors) and their actions. Our transformer-based model learns to separate actor and action features effectively by employing supervised contrastive losses alongside standard cross-entropy loss to ensure proper feature separation.

L'objectif de cette thèse est d'améliorer les représentations des images et des vidéos dans le but d'obtenir une reconnaissance visuelle accrue, tant pour des entités spécifiques que pour des catégories plus génériques. Les contributions de cette thèse portent, pour l'essentiel, sur des méthodes de description du contenu visuel. Nous proposons des méthodes pour la recherche d'image par le contenu ou par des requêtes textuelles, ainsi que des méthodes pour la reconnaissance et la localisation d'action dans des vidéos. En recherche d'image, les contributions se fondent sur des méthodes `a base de plongements de Hamming. Tout d'abord, une méthode de comparaison asymétrique vecteur-`a-code est proposée pour améliorer la méthode originale, symétrique et utilisant une comparaison code-`a-code. Une méthode de classification fondée sur l'appariement de descripteurs locaux est ensuite proposée. Elle s'appuie sur une classification opérée dans un espace de similarités associées au plongement de Hamming. En reconnaissance d'action, les contributions portent essentiellement sur des meilleures manières d'exploiter et de représenter le mouvement. Finalement, une méthode de localisation est proposée. Elle utilise une partition de la vidéo en super-voxels, qui permet d'effectuer un échantillonnage 2D+t de suites de boîtes englobantes autour de zones spatio-temporelles d'intérêt. Elle s'appuie en particulier sur un critère de similarité associé au mouvement. Toutes les méthodes proposées sont évaluées sur des jeux de données publics. Ces expériences montrent que les méthodes proposées dans cette thèse améliorent l'état de l'art au moment de leur publication. / The subject of this thesis is about image and video representations for visual recognition. This thesis first focuses on image search, both for image and textual queries, and then considers the classification and the localization of actions in videos. In image retrieval, images similar to the query image are retrieved from a large dataset. On this front, we propose an asymmetric version of the Hamming Embedding method, where the comparison of query and database descriptors relies on a vector-to-binary code comparison. For image classification, where the task is to identify if an image contains any instance of the queried category, we propose a novel approach based on a match kernel between images, more specifically based on Hamming Embedding similarity. We also present an effective variant of the SIFT descriptor, which leads to a better classification accuracy. Action classification is improved by several methods to better employ the motion inherent to videos. This is done by dominant motion compensation, and by introducing a novel descriptor based on kinematic features of the visual flow. The last contribution is devoted to action localization, whose objective is to determine where and when the action of interest appears in the video. A selective sampling strategy produces 2D+t sequences of bounding boxes, which drastically reduces the candidate locations. The method advantageously exploits a criterion that takes in account how motion related to actions deviates from the background motion. We thoroughly evaluated all the proposed methods on real world images and videos from challenging benchmarks. Our methods outperform the previously published related state of the art and remains competitive with the subsequently proposed methods.

As imaging systems become ubiquitous, the ability to recognize human actions is becoming increasingly important. Just as in the object detection and recognition literature, action recognition can be roughly divided into classification tasks, where the goal is to classify a video according to the action depicted in the video, and detection tasks, where the goal is to detect and localize a human performing a particular action. A growing literature is demonstrating the benefits of localizing discriminative sub-regions of images and videos when performing recognition tasks. In this thesis, we address the action detection and recognition problems. Action detection in video is a particularly difficult problem because actions must not only be recognized correctly, but must also be localized in the 3D spatio-temporal volume. We introduce a technique that transforms the 3D localization problem into a series of 2D detection tasks. This is accomplished by dividing the video into overlapping segments, then representing each segment with a 2D video projection. The advantage of the 2D projection is that it makes it convenient to apply the best techniques from object detection to the action detection problem. We also introduce a novel, straightforward method for searching the 2D projections to localize actions, termed TwoPoint Subwindow Search (TPSS). Finally, we show how to connect the local detections in time using a chaining algorithm to identify the entire extent of the action. Our experiments show that video projection outperforms the latest results on action detection in a direct comparison. Second, we present a probabilistic model learning to identify discriminative regions in videos from weakly-supervised data where each video clip is only assigned a label describing what action is present in the frame or clip. While our first system requires every action to be manually outlined in every frame of the video, this second system only requires that the video be given a single highlevel tag. From this data, the system is able to identify discriminative regions that correspond well iii to the regions containing the actual actions. Our experiments on both the MSR Action Dataset II and UCF Sports Dataset show that the localizations produced by this weakly supervised system are comparable in quality to localizations produced by systems that require each frame to be manually annotated. This system is able to detect actions in both 1) non-temporally segmented action videos and 2) recognition tasks where a single label is assigned to the clip. We also demonstrate the action recognition performance of our method on two complex datasets, i.e. HMDB and UCF101. Third, we extend our weakly-supervised framework by replacing the recognition stage with a twostage neural network and apply dropout for preventing overfitting of the parameters on the training data. Dropout technique has been recently introduced to prevent overfitting of the parameters in deep neural networks and it has been applied successfully to object recognition problem. To our knowledge, this is the first system using dropout for action recognition problem. We demonstrate that using dropout improves the action recognition accuracies on HMDB and UCF101 datasets.

Vidéo d'interprétation et de compréhension est l'un des objectifs de recherche à long terme dans la vision par ordinateur. Vidéos réalistes tels que les films présentent une variété de problèmes difficiles d'apprentissage machine, telles que la classification d'action / récupération d'action, de suivi humaines, la classification interaction homme / objet, etc Récemment robustes descripteurs visuels pour la classification vidéo ont été développés, et ont montré qu'il est possible d'apprendre classificateurs visuels réalistes des paramètres difficile. Toutefois, afin de déployer des systèmes de reconnaissance visuelle à grande échelle dans la pratique, il devient important d'aborder l'évolutivité des techniques. L'objectif principal est cette thèse est de développer des méthodes évolutives pour l'analyse de contenu vidéo (par exemple pour le classement ou la classification). / Video interpretation and understanding is one of the long-term research goals in computer vision. Realistic videos such as movies present a variety of challenging machine learning problems, such as action classification/action retrieval, human tracking, human/object interaction classification, etc. Recently robust visual descriptors for video classification have been developed, and have shown that it is possible to learn visual classifiers in realistic difficult settings. However, in order to deploy visual recognition systems on large-scale in practice it becomes important to address the scalability of the techniques. The main goal is this thesis is to develop scalable methods for video content analysis (eg for ranking, or classification).

No description available.