Abstract: The regression-based 2D human pose estimation methods directly predict the coordinates of human keypoints.The transformer can effectively establish the relationship between human body parts,and its application significantly improves the accuracy of the regression-based methods.However,related methods have the following two problems:1)In the cross-attention module,for different images,the fixed query can not properly focus on different keypoint regions,which leads to distraction.2)They directly learn the labeled keypoint coordinates and overfit annotations.In this paper,a pose estimation model based on adaptive prediction is proposed to solve these two problems.For the first problem,the model adaptively predicts the region of attention of the query and directs the attention to that region.For the second problem,the model adaptively predicts the probability distribution of keypoint appearing in every position,and alleviates the model's overfitting to annotations by means of soft prediction.Experiments on the MS-COCO dataset show that the model improves the accuracy of the baseline method by 2.8% and improves the highest accuracy of related methods by 0.2%.

Key words: 2D human pose estimation, Regression-based, Adaptive, Region of attention, Probability distribution