Beyond Perspectives: Enhancing Pose Estimation via Viewpoint Transformation

A number of approaches have been proposed to address the occlusion and truncation issues in pose estimation, but they typically require additional devices or specific recording environments. In this paper, we propose a novel pose estimation mechanism that utilizes GAN-based viewpoint transformation. Our mechanism complements missing pose information without requiring additional devices or a pre-aligned camera setups. It achieves this by transforming a supplementary viewpoint video to the target viewpoint video utilizing GAN and using it to complement missing keypoints in the target viewpoint video. The evaluation results confirm that our mechanism effectively complements missing pose information and provides reliable viewpoint transformation performance.