Toward Precise Point Location in Mouse Pose Estimation by Multi-scale Supervised Network
Junbiao Pang, Shuhong Wan · 2021 China Automation Congress (CAC) · 2021
Mouse pose estimation has been used in animal behavior analysis that plays an important role in biological neuroscience and behavioral science. Computer vision community mainly focuses on human pose estimation. Human pose estimation locates joints by regressing the Gaussian heatmaps of joints. Compared with the human pose estimation task, the mouse pose estimation requires the precise location of key points. A naive solution is that mouse pose estimation borrows the network from the human pose etstimation with the smaller Guassian heatmaps. However, smaller Gaussian heatmaps empirically make the model difficult to be trained and even not be converged, leading to an inferior accuracy. In this paper, we propose a muti-scale supervised network by adding the spatial constraints into the state-of-the- art the top-down to down-top structure. Concretely, the small Gaussian heatmaps are added as supervision informations at the each scale of the down-top path. With the multi-scale supervison, the resulting features of the down-top scales progressively locate the joints and make training process fast to converge. Besides, a multi-stage fusion is proposed to makes that the multi-scale supervison at the down-top path could transfer into the top- down one from the next stage. Two datasets, a benchmark dataset and a dataset published by ourself, are used to demonstrate the effectiveness of methods.