Abstract
While being the de facto standard coordinate representation in human poseestimation, heatmap is never systematically investigated in the literature, toour best knowledge. This work fills this gap by studying the coordinaterepresentation with a particular focus on the heatmap. Interestingly, we foundthat the process of decoding the predicted heatmaps into the final jointcoordinates in the original image space is surprisingly significant for humanpose estimation performance, which nevertheless was not recognised before. Inlight of the discovered importance, we further probe the design limitations ofthe standard coordinate decoding method widely used by existing methods, andpropose a more principled distribution-aware decoding method. Meanwhile, weimprove the standard coordinate encoding process (i.e. transformingground-truth coordinates to heatmaps) by generating accurate heatmapdistributions for unbiased model training. Taking the two together, weformulate a novel Distribution-Aware coordinate Representation of Keypoint(DARK) method. Serving as a model-agnostic plug-in, DARK significantly improvesthe performance of a variety of state-of-the-art human pose estimation models.Extensive experiments show that DARK yields the best results on two commonbenchmarks, MPII and COCO, consistently validating the usefulness andeffectiveness of our novel coordinate representation idea.