FasterPose: A Faster Simple Baseline for Human Pose Estimation

The performance of human pose estimation depends on the spatial accuracy keypoint localization. Most existing methods pursue through learning high-resolution (HR) representation from input images. By experimental analysis, we find that HR leads to a sharp increase computational cost, while improvement remains marginal compared with low-resolution (LR) representation. In this article, propose design paradigm for cost-effective network LR efficient estimation, named FasterPose. Whereas largely shrinks model complexity, how effectively train respect is concomitant challenge. We study training behavior FasterPose and formulate novel regressive cross-entropy (RCE) loss function accelerating convergence promoting accuracy. RCE generalizes ordinary binary supervision continuous range, thus able benefit sigmoid function. doing so, output heatmap can be inferred features without accuracy, cost size has been significantly reduced. Compared previously dominant our method reduces 58% FLOPs simultaneously gains 1.3% Extensive experiments show yields promising results common benchmarks, i.e., COCO MPII, consistently validating effectiveness efficiency practical utilization, especially low-latency low-energy-budget applications in non-GPU scenarios.