How to Train Your Robust Human Pose Estimator: Pay Attention to the Constraint Cue
share
Both appearance cue and constraint cue are important in human pose estimation. However, the widely used response map supervision has the tendency to overfit the appearance cue and overlook the constraint cue. In this paper, we propose occlusion augmentation with customized training schedules to tackle this dilemma. Specifically, we implicitly force the neural network focus on the constraint cue by dropping appearance information within keypoint-aware strategy. Besides, a two-steps schedule is designed to deal with the information shortage in early training process, which effectively exploits the potential of the proposed occlusion augmentation. In experiments, as a model-agnostic approach, occlusion augmentation consistently promotes most SOTAs with different input sizes, frameworks, backbones, training and test sets. For HRNet within W32-256x192 and W48plus-384x288 configurations, occlusion augmentation obtains gains by 0.6 AP (75.6 to 76.2) and 0.7 AP (76.8 to 77.5) on COCO test-dev set, respectively. HRNet-W48plus-384x288 equipped with extra training data and occlusion augmentation achieves 78.7 AP. Furthermore, the proposed occlusion augmentation makes a remarkable improvement on more challenging CrowdPose dataset. The source code will be publicly available for further research in https://github.com/HuangJunJie2017/UDP-Pose.
READ FULL TEXT