SimpleCut: A simple and strong 2D model for multi-person pose estimation

摘要

This article proposes a simple and efficient multi-person pose estimation model which follows a bottom-up approach and is based on a few deconvolutional layers added on a U-net lookalike ResNet featuremap. SimpleCut contains four independent modules: joints module, coordinates (coords) module, main-joint pairing module, and other-joints pairing module. The joints module builds a score on joints of each individual on the image, whereas the coords module encodes the location of those joints, and both the pairing modules generate image-conditioned pairing of the joints on a small scale. The pairing modules help set up the proposals into a variable number of consistent body part configurations by an optimization strategy that efficiently brings significant speed-up factors. We demonstrated that simultaneously inferring these bottom-up representations of detection and association encode global context sufficiently well to allow a greedy parse to attain high-quality results with low computational cost. SimpleCut evaluated on three publicly available large-scale dataset benchmarks such as ms-coco, lspet, and mpii human pose dataset.