Visual representation learning by exploring spatio-temporal consistency

Video representation learning is a fundamental area of research in computer vision. It focuses on developing methods and models to encode video data into definitive and discriminative representations that can be effectively utilized in downstream tasks, such as video classification, action recognition, video retrieval and video captioning. At the core of it, video representation learning seeks to understand scene, actor and actor's relationship with their surroundings, e.g. objects, and scene elements. In order to learn the transition of visual content over time, a model needs to capture temporal and spatial information inherently encoded in video sequences and to extract spatiotemporal representations that can be useful for downstream applications. This thesis addresses the problem of self-supervised video representation learning focused on motion features, aiming to capture features from foreground motion with reduced reliance on background bias. Recent successful methods often employ instance discrimination approaches, which entail heavy computation and may lead to inefficient and exhaustive pretraining. Moreover, proposed methods in this thesis do not require any additional computation for motion features, such as optical flow, H264 codec features or any motion descriptor. Although several works in literature incorporate two-stream networks to incorporate motion learning, this thesis work seeks for single network solutions for learning spatiotemporal features. To this end, we utilize the augmentation technique MAC: Mask-Augmentation TeChnique. MAC blends foreground motion using frame-difference-based masks and sets up a pretext task to recognize the applied transformation. By incorporating a game of predicting the correct blending multiplier at the pretraining stage, our model is compelled to encode motion-based features, which are then successfully transferred to downstream tasks such as action recognition and video retrieval. Moreover, we expand our approach within a joint contrastive framework and integrate additional tasks in the spatial and temporal domains to further enhance representation capabilities. We present experimental results on action recognition and video retrieval tasks to demonstrate that our method achieves superior performance on the UCF-101, HMDB51 and Diving-48 datasets under low-resource settings and competitive results with instance discrimination methods under costly computation settings. We carefully design ablation experiments to analyze learning behavior of proposed methods and contribution of each component. Lastly, we present qualitative results to further illustrate the benefits of presented methods. We anticipate that our augmentation technique, along with the associated pretext and contrastive learning objectives, will lay the groundwork for future advancements in self-supervised video representation learning.