Human Pose Estimation via Dynamic Information Transfer

This paper presents a multi-task learning framework, called the dynamic information transfer network (DITN). We mainly focused on improving the pose estimation with the spatial relationship of the adjacent joints. To benefit from the explicit structural knowledge, we constructed two branches with a shared backbone to localize the human joints and bones, respectively. Since related tasks share a high-level representation, we leveraged the bone information to refine the joint localization via dynamic information transfer. In detail, we extracted the dynamic parameters from the bone branch and used them to make the network learn constraint relationships via dynamic convolution. Moreover, attention blocks were added after the information transfer to balance the information across different granularity levels and induce the network to focus on the informative regions. The experimental results demonstrated the effectiveness of the DITN, which achieved 90.8% PCKh@0.5 on MPII and 75.0% AP on COCO. The qualitative results on the MPII and COCO datasets showed that the DITN achieved better performance, especially on heavily occluded or easily confusable joint localization.