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 pose estimation with spatial relationship of adjacent joints. To benefit from explicit structural knowledge, we constructed two branches shared backbone to localize human joints and bones, respectively. Since related tasks share high-level representation, leveraged bone refine joint localization via transfer. In detail, extracted parameters branch used them make learn constraint relationships convolution. Moreover, attention blocks were added after balance across different granularity levels induce focus informative regions. The experimental results demonstrated effectiveness DITN, which achieved 90.8% [email protected] MPII 75.0% AP COCO. qualitative COCO datasets showed that DITN better performance, especially heavily occluded or easily confusable localization.