Decoupled Depth-guided Coordinate-aware Network for 3D Human Pose and Shape Estimation

3D human pose and shape estimation is a fundamental task in human-centric computer vision. However, current end-to-end approaches face significant challenges, among which depth ambiguity remains a critical issue. To address this, we propose a Decoupled Depth-guided Coordinate-aware Network. Our approach designs a novel average camera-depth estimator to capture explicit depth information. Furthermore, we introduce an innovative coordinate-aware attention module, which can encode the coupled pose feature and decouple it into 3D coordinate features. Such a decoupling operation ensures that the depth information provides explicit guidance only for the necessary coordinate features. A multidimensional parallel regressor is proposed to further facilitate the decoupled coordinate features and improve the performance of estimation. Experimental results demonstrate that the proposed method not only effectively avoids depth ambiguity but also achieves state-of-the-art performance on Human3.6M and 3DPW datasets.