View-Invariant, Occlusion-Robust Probabilistic Embedding for Human Pose

Recognition of human poses and actions is crucial for autonomous systems to interact smoothly with people. However, cameras generally capture in 2D as images videos, which can have significant appearance variations across viewpoints that make the recognition tasks challenging. To address this, we explore recognizing similarity 3D body from information, has not been well-studied existing works. Here, propose an approach learning a compact view-invariant embedding space joint keypoints, without explicitly predicting poses. Input ambiguities projection occlusion are difficult represent through deterministic mapping, therefore adopt probabilistic formulation our space. Experimental results show model achieves higher accuracy when retrieving similar different camera views, comparison pose estimation models. We also by training simple temporal model, achieve superior performance on sequence retrieval largely reduce dimension stacking frame-based embeddings efficient large-scale retrieval. Furthermore, order enable work partially visible input, further investigate keypoint augmentation strategies during training. demonstrate these augmentations significantly improve partial input Results action video alignment using any additional competitive relative other models specifically trained each task.