3D Human Pose Estimation Based On Geometric Model

3D human pose estimation is one of the most important tasks in computer vision,as well as the core technologies of human-computer interaction.The impressive progress of pose estimation,especially of 2D pose estimation and object detection,reveals possibilities in 3D pose estimation.There are two main aspects of geometric models in 3D computer vision problem,one of them is the camera-projection geometry,and the other is the geometric representation of objects,the human body.As to the camera-projection geometries,most of the current datasets are vulnerable to the limited camera views and easy to be over-fitting to existed camera configurations.As to the objects,human body in 3D pose estimation shall cover not only the poses,but shapes,materials,reflections,and textures as well.Therefore,camera-view generalization oriented data augmentation and complex geometric model for human body with both pose and shape encoded should be considered.To this end,we propose a complete framework for virtual camera augmentation and an animation dance motion dataset for camera generalization training and evaluation.Furthermore,we design an end-to-end 3D human pose estimation and shape recovery pipeline,based on a kind of ellipsoidrepresentation for human body,named as Ellip Body.With a powerful partlevel differentiable renderer(Part DR)and pixel-wise shape hints from body part segmentation,our pipeline achieves the state-of-the-art performance compared to other regression or optimization based approaches in 3D human pose estimation.Ellip Body can also perform as a flexible and light-weight intermediate representation pointed to other parametric geometry models.