DoubleHigherNet: coarse-to-fine precise heatmap bottom-up dynamic pose computer intelligent estimation

Accurate keypoint positioning is necessary for bottom-up multi-person pose estimation methods to handle scale variation and crowdedness. In this paper, we present DoubleHigherNet: a novel network learning scale-aware and precise heatmap representation for bottom-up process using double high-resolution feature pyramids and coarse-to-fine training. The two feature pyramids in DoubleHigherNet consists of 1/4 resolution feature and higher-resolution (1/2) maps generated by attention fusion blocks and transposed convolutions. Benefited by the training strategy, muti-resoltion and coarse-fine heatmap aggregation, the proposed approach is able to predict keypoints more accurately so as to perform better on difficult crowded scenes. DoubleHigherNetw32 achieves competitive result on CrowdPose-test, surpassing all the top-down methods and bottom-up SOTA HigherHRNet-w32 (which possesses similar number of params with DoubleHigherNet-w32).