Joint Model for Human Body Part Instance Segmentation and DensePose Estimation

Abstract

Human body part instance segmentation and human body densepose estimation are new frontier research problems of great significance, they have high practical application and research value. We will propose a new unified algorithm framework that can train these two tasks at the same time. First, we will build a joint model for these two tasks based on the basic framework of MASK-RCNN; In addition, we will propose a new lightweight attention mechanism module LAM, which reduces the amount of parameters and calculations more than ten times compared with the GCE module in Parsing-RCNN; We combine LAM and the improved dense semantic analysis module based on DenseASPP to form deep and dense feature information. The extraction module LADCEM is lighter than the GCE module, but can bring more accuracy improvements; Then we will use anchor freed FCOS of one-stage which is full convolution detection module for the first time to replace the two-stage anchor based RPN, which can significantly improve the accuracy of human detection, both subtasks are greatly improved. In the end, the MobileNet-V3 network is used as the basic Backbone, the overall parameter amount is about 10% of the Backbone based on Rensenet50, making it easier to deploy on mobile device using our model.