Fusion scale adaptive multi-person pose estimation
Wen Wang, QiongYan Wu, Zhengpeng Zhao, ChenYang Qiu, Yuanyuan Pu, Xiaolong Liu · 2023
Currently, the mainstream top-down multi-person pose estimation algorithms based on convolutional neural networks have achieved high estimation accuracy, but in the standard convolutional neural networks, the perceptual field size of artificial neurons in each layer is designed to be fixed, lacking consideration of the multi-scale problem. To solve this problem, we propose a scale-adaptive multi-person pose estimation method based on the SKNet network introduced in the feature extraction phase of the pose estimation network, which improves the model's ability to extract important features at different scales by weighted fusion of features under different scale branches. DO-Conv is used in the pose estimation network to replace the original two-dimensional convolution, so that it can better capture the important features in the input image, thus further improving the algorithm performance. In addition, a better human detector can improve the performance of the subsequent single person pose estimation network to a certain extent, so the latest YOLOv7 target detector is used in this paper as the human detector in the multi-person pose estimation network. The proposed algorithm is based on the current mainstream multi-person pose estimation algorithms for comparison experiments and shows better performance than the original network itself in the validation experiments on the COCO dataset and the Halpe dataset.