Human Pose Estimation using SimCC and Swin Transformer

Abstract

2D Human Pose Estimation is an important task in computer vision. In recent years, methods using deep learning for human pose estimation have been proposed one after another and achieved good results. Among existing models, the built-in attention layer in Transformer enables the model to effectively capture long-range relationships and also reveal the dependencies on which predicted key points depend. SimCC formulates keypoint localization as a classification problem, dividing the horizontal and vertical axes into equal-width numbered bins, and discretizing continuous coordinates into integer bin labels. We propose a new model that combines the Swin Transformer training model to predict the bin where the key points are located, so as to achieve the purpose of predicting key points. This method can achieve better results than other models and can achieve supixel positioning accuracy and low quantization error.