Nearby-person Occlusion Data Augmentation for Human Pose Estimation with Non-extra Annotations

Human pose estimation has made significant progress with deep learning techniques, while the estimation of occlusion keypoints is still an unsolved problem. One important reason comes from the insufficiency of the existing benchmark datasets, such as imbalanced body keypoints annotation and lack of occluded training samples. To address this problem, we propose Nearby-person Occlusion Data Augmentation (NODA), a method that provides synthetic nearby-person occlusion images by only utilizing existing annotations. First, we generate rough mask of human bodies with keypoints annotation to build a foreground human body pool. Then, one foreground human body crop is randomly sampled and properly placed over the training human body to synthesis nearby-person occlusion training images. The proposed data augmentation method is easy to implement and deploy to any other methods. Extensive experimental results on MPII benchmark demonstrate the effectiveness of our method with Simple and HRNet as the backbone models. Especially on easily-confusable joints, our method makes significant improvement.