Offset-based Disentangled Representation for Efficient Human Pose Estimation

2D heatmap-based human pose estimation has exhibited remarkable performance. However, constrained by quantization error, heatmap-based methods heavily rely on high-resolution heatmaps and intricate post-processing to enhance detection accuracy, thereby incurring substantial computational costs. To pursue more effective keypoint representation, we propose a novel scheme named Offset-based Disentangled Representation (ODR). ODR conducts coordinate classification and offset prediction simultaneously using 1D vectors and aggregates their outputs to precisely pinpoint keypoint positions, thus freeing from dependence on high-resolution heatmaps. To eliminate the impact of long-range offsets, we propose a scale-aware eraser that generates noticeable intervals based on the relative scale of different keypoints, directing the regression task to focus on short-range offsets. By doing so, upsampling layers are no longer necessary, enabling a more concise and effective architecture for human pose estimation. Extensive experiments conducted over COCO and MPII datasets validate the superiority of ODR over counterparts based on 2D or 1D heatmaps. Our source codes are available at the link.