SE-PifPaf: A Multi-Person Human Pose Estimation Methond for Complex Scenes
Tianbao Wang, Caixia Meng, Lei Shi, Yufei Gao, Qingxian Wang, Lin Wei · 2023
Images in complex scenes are prone to cross-obscuration of people, which brings great challenges to multi-person human pose estimation. Previous multi-person human pose estimation algorithms suffer from limited accuracy and human skeleton loss in complex scenes. To solve the above problems, a novel pose estimation optimization algorithm SE-PifPaf is proposed. Firstly, the grouped convolution network is employed to mine semantic features relevant to the human pose estimation task. Secondly, the attention mechanism is applied to realize the interaction among channels, obtaining the weight of each channel on the pose estimation task in complex scenes. Experimental results show that SE-PifPaf achieves an accuracy of 72.3% on COCO 2017 dataset. More importantly, it improves the accuracy of pose estimation by 1.1% in complex scenes on CrowdPose dataset.