Local feature enhancement for robust 2D multi-person pose estimation via pose refinement network

Accurate 2D multi-person pose estimation remains challenging due to issues such as occlusion, missing body parts, and low resolution, particularly in complex backgrounds. This paper proposes an refinement network for multi-person pose estimation through the complementary fusion of extended local receptive fields and contextual information. The proposed cascaded dilated convolution module (DCM) expands the local receptive field through geometric perception, addressing the issue of feature ambiguity in low-resolution and small-scale human bodies. Simultaneously, a hybrid self-attention module (HSM) is introduced to integrate the semantic relevance of joints and precise spatial location information by parallelly combining convolutional self-attention (CSA) and coordinate attention (CA). This optimizes localization through semantic association, not only reducing background interference but also resolving the problem of overlapping joints in multi-person scenarios. Consequently, the network framework achieves an effective balance between the accuracy of human feature extraction at different scales and computational speed. Extensive experiments conducted on the MS COCO and CrowdPose datasets demonstrate that the proposed network architecture outperforms comparable methods, exhibiting superior robustness and computational performance in high-density crowd scenes, uneven lighting conditions, and complex texture scenarios. The related code and models are available at https://github.com/Twl-GZ/Human-pose .