A Feature Fusion Based Dense Estimation Model for 3D Human Pose

With the development of technology, especially the rise of some practical application technology such as AR, VR, and human-computer interaction, 3D human pose estimation has shown more and more extensive prospects. However, in existing 3D human pose estimation methods, there are still some issues. For example, the lack of multi-scale and diversity in the analysis of human posture leads to insufficient refinement in fitting and segmentation of body parts, and eventually causing unsatisfactory performance. In this paper, we propose a feature-fusion-based dense estimation model (FFD model) for 3D human pose estimation to solve the above problems. Specifically, we design a feature fusion module obtaining multi-scale feature information, which not only captures the overall information about the human posture, but also retains the position information and detailed information of the human body part. Furthermore, we use a dense feature extraction backbone network to expand the diversity of receptive fields, thus making the human pose estimation more refined and precise. Experiments on the COCO dataset show the superiority of FFD model compared with the baseline model.