Human behavior can be described at multiple levels. At the lowest level, we observe the 3D pose of the body over time. Poses can be organized into primitives that capture coordinated activity of different body parts. These further form more complex "actions" or "behaviors". Finally, underlying all of the above are the goals, motives, and emotions of the person; that is, the cause of the movement. Our ultimate goal is to extract this high-level causal information from video.
Human behavior and action analysis based on visual data is one of the most essential tasks in computer vision, the insights and results of which can lead to a large number of applications in diverse scenarios, such as patient healthcare, assistive living, animation, video retrieval and beyond. Therefore, we conduct such Action and behavior project, towards thorough understanding of human behaviors and providing effective solutions to various practical issues. Referring to how human beings understand behaviors of others, our research is performed from two perspectives:
Low-level understanding. Humans can readily differentiate biological motion from non-biological motion. They can do this from sparse visual cues like moving dots and without any explicit supervision. In this spirit, we perform behavior analysis at a low-level using a novel dynamic clustering algorithm that groups actions in an online fashion [ ]. As a building block, dynamic clustering is employed in a computational pipeline, where low-level visual cues are aggregated to high-level action patterns via temporal pooling. Our experiments show that such hierarchical dynamic clustering scheme is reliable, generic for diverse input features and fast.
High-level understanding. Here we relate low-level behavior to high-level concepts by identifying individual actors and synthesizing natural-language descriptions of their actions and interactions. We do so using weak supervision provided by scripts associated with the video. As a first attempt, we generate descriptions with grounded and co-referenced people [ ]. Specifically, we first learn to localize characters by relating their visual appearance to mentions in the descriptions via a semi-supervised approach. We then provide this (noisy) supervision to a description model, which greatly improves its performance. Our proposed description model improves over prior work w.r.t. generated description quality and additionally provides grounding and local co-reference resolution.
Ongoing work leverages richer models of the human and their motion as well as richer models of the scene and objects in it.