Bias-Compensated Integral Regression for Human Pose Estimation

In human and hand pose estimation, heatmaps are a crucial intermediate representation for body or keypoint. Two popular methods to decode the heatmap into final joint coordinate via an argmax, as done in detection, softmax expectation, integral regression. Integral regression is learnable end-to-end, but has lower accuracy than detection. This paper uncovers induced bias from that results combining expectation operation. often forces network learn degenerately localized heatmaps, obscuring keypoint's true underlying distribution leads accuracies. Training-wise, by investigating gradients of regression, we show implicit guidance update makes it slower converge To counter above two limitations, propose Bias Compensated Regression (BCIR), regression-based framework compensates bias. BCIR also incorporates Gaussian prior loss speed up training improve prediction accuracy. Experimental on both benchmarks faster train more accurate original making competitive with state-of-the-art detection methods.