Abstract: Recently, the leading performance of human pose estimation is dominated by heatmap-based methodsalgorithms. Heatmap decoding (i.e. transforming heatmaps to coordinates) is a fundamental link of those algorithms. Previous heatmap decoding methods generally neglect the effects of systematic errors. Therefore, an error-compensation-based heatmap decoding method is proposed in this work, which estimates an error compensation factor during training and then compensates both systematic and random errors during inferring. While being a fundamental link of those methods, the heatmap decoding process (i.e. transforming heatmaps to coordinates) receives limited investigations. This work fills the gap by studying the heatmap decoding with a particular focus on errors. Extensive experiments on different network architectures, input resolutions, evaluation metrics and datasets validatedwas tested. The results show that compared with previous state-of-the-art method, significant accuracy gain can be achieved with thisthe proposd method. We, for the first time, revealed that heatmap-based methods suffer from significant errors, which nevertheless was ignored before. To tackle this issue, we define an optimal error compensation factor Δ_opt to describe the error properties of a network and use Δ_opt to compensates errors. Our method reduces the systematic and random errors in one shot with negligible extra computation. Extensive experiments with different network architectures, input sizes, evaluation metrics, and datasets have shown that the proposed method achieves significant accuracy gains over the state-of-the-art heatmap decoding method. Specifically, the HRNet-W48-256×192 model is improved by 2.9 AP on the COCO dataset and the PCKh0.1 of the ResNet-152-256×256 model is improved by 7.8% on the MPII dataset. Besides, unlike the existing methods, our the proposed method is smoothing-free and is 2 times faster (1.4 vs 3.0 ms/f) than the state-of-the-art. ThisThe work brought applicable values to developing fast and accurate human pose estimation .