Human pose estimation for low-resolution image using 1-D heatmaps and offset regression

Running a reliable network on resource-limited platforms for a low-resolution image is a great challenge for heatmap-based human pose estimation (HPE). Scale mismatch between the input image and heatmaps and the intrinsic quantization effect induced by the 'argmax' function hinder the performance of heatmap-based human pose estimation for low-resolution image. In this paper, we propose a coordinate-decoupled and offset-revised module (CDORM) to tackle these challenges. The proposed CDORM uses two coordinate-decoupled 1-D heatmaps to supervise the regression process of determining the horizontal and vertical locations of human joints, and employs offset regressing to alleviate the effect of quantization. The CDORM can be integrated with any current heatmap-based HPE network without increasing the size of network significantly. Experimental results on the COCO and MPII datasets show that CDORM helps heatmap-based regression approaches obtain high estimation accuracy from the low-resolution image and only slightly increases the size and runtime of the network.