A Compact and Powerful Single-Stage Network for Multi-Person Pose Estimation

Multi-person pose estimation generally follows top-down and bottom-up paradigms. The paradigm detects all human boxes then performs single-person on each ROI. locates identity-free keypoints groups them into individuals. Both of use an extra stage to build the relationship between instance corresponding (e.g., detection in a manner or grouping process manner). leads high computation cost redundant two-stage pipeline. To address above issue, we introduce fine-grained body representation method. Concretely, is divided several local parts part represented by adaptive point. novel able sufficiently encode diverse information effectively model single-forward pass. With proposed representation, further compact single-stage multi-person regression network, called AdaptivePose++, which extended version AAAI-22 paper AdaptivePose. During inference, our network only needs single-step decode operation estimate without complex post-processes refinements. Without any bells whistles, achieve most competitive performance representative 2D benchmarks MS COCO CrowdPose terms accuracy speed. In particular, AdaptivePose++ outperforms state-of-the-art SWAHR-W48 CenterGroup-W48 3.2 AP 1.4 mini-val with faster inference Furthermore, outstanding 3D datasets MuCo-3DHP MuPoTS-3D demonstrates its effectiveness generalizability scenes.