Lightweight Human Pose Estimation Based on Heatmap Weighted Loss Function

research on human pose estimation often focuses on using complex structures to improve task accuracy, while overlooking resource consumption and inference speed during actual deployment. Based on the LitePose pose estimation architecture, this paper proposes a lightweight bottom-up pose estimation model, WLitePose, designed to better handle complex scenes. Specifically, to address the limitations of the MSE loss function, a heatmap weighted loss function is proposed to enable the model to focus more on the areas surrounding the true keypoint locations during training. To enhance the model's ability to handle variations in human scale, a lightweight deconvolution module is used after the main architecture to generate higher-resolution heatmaps. During the inference phase, heatmaps of different resolutions are aggregated. Additionally, the DFC-bottleneck block is proposed to enhance the backbone network's ability to capture long-range dependence between different spatial pixels. Experimental results on the COCO and CrowdPose datasets demonstrate that the proposed model achieves a good balance between task accuracy and computational complexity.