LAR-Pose: Lightweight human pose estimation with adaptive regression loss

Abstract

In this paper, LAR-Pose, a lightweight, high-resolution network for human pose estimation driven by adaptive regression loss is proposed and experimentally demonstrated based on MS COCO and MPII. The architecture of the LAR-Pose comprises two main components. One is a lightweight high-resolution backbone network, which utilizes a parallel high-resolution architecture with conditional channel weighting block to reduce the model size and computational complexity. The other is a dynamic residual refinement network, which calculates residuals from pseudo-heatmaps and scaling factors, improving training concentration for consistent distribution estimation, rather than predicting coordinates or heatmaps directly. Specific coordinates are derived through integral heatmap regression, effectively minimizing quantization errors. Our adaptive regression loss, which uses a flow model to fit the distribution of residuals in real-time, provides more sensitive parameter feedback than conventional heatmap loss, ensuring differentiability and continuity during backpropagation while enhancing performance. With a relatively small parameter scale, LAR-Pose achieves an AP of 73.5 on MS COCO and a PCKh of 90.9 on MPII, while the results outperform most advanced small networks and approach the performance of large networks.