Research On Human Body Part Instance Segmentaition And Human Body Densepose Estimation Method

Human body instance segmentation and human body densepose estimation algorithms are key technologies in the field of computer vision.The segmentation of human body parts has important application value and significance for scenarios such as security monitoring,virtual reality,clothing recognition,and retrieval.Human body densepose estimation needs to build a dense mapping between the human body 3D model and 2D image,which can help enhance people’s understanding of twodimensional and three-dimensional relationships.The current research on both tasks is separate and independent,so the training efficiency is relatively low,the connection between them cannot be well explored,and information between them cannot be shared.Based on the above research background,this paper mainly studies the construction of a unified joint model and the relationship dependence network between the two.It is mainly divided into the following two contents.This article first builds a multi-task joint training framework,which simplifies the complexity of multi-task training,realizes the parallel processing of multi-tasks,and improves the performance of multi-task algorithms.In addition,in order to reduce the number of model parameters of the framework,a new type of lightweight attention mechanism module LAM is proposed,which reduces the number of parameters and calculations by tens of times compared with the GCE module in Parsing-RCNN.In order to make the analysis module of human body instance segmentation lighter,this paper combines LAM and an improved dense semantic analysis module based on DenseASPP to extract in-depth information,which saves 40%of the parameters and improves the accuracy by 2%.In order to improve the accuracy of human detection,this paper uses the FCOS one-stage Anchor-free full convolution detection module to replace the two-stage Anchor-based RPN,which has improved accuracy in both subtasks.This paper then discusses the important influence of the human body part segmentation task on the human body dense pose estimation algorithm,as well as their mutual enhancement.The proposed dependency network can further associate the human body instance part segmentation with the human body dense pose estimation algorithm,which can improve the performance of the algorithm.A data processing method that can transfer different human body instance segmentation data sets is proposed.In the process of enhancing the performance of the instance segmentation algorithm,the performance of the human body intensive pose estimation algorithm can be improved at the same time.