Dual Stream Networks for 3d Human Pose and Shape Estimation

This work addresses the problem of 3D human pose and shape estimation from monocular video. While significant progress has been made in this domain, most existing approaches focus solely on forward human motion and neglect the reverse process. To overcome this oversight, we propose dual stream networks consisting of decoders that analyze both forward and reverse human motion. The network outputs are fused using residual connectors, which effectively integrate the predictions. Exploiting temporal transformers further enhance ac-curacy. Our method achieves state-of-the-art results on well-established datasets in terms of both 3D error and temporal consistency metrics, demonstrating its robustness and effectiveness. Furthermore, our network operates at over 150 FPS, making it suitable for real-time inference.