Multi-person Pose Estimation with Object Occlusion Information

Multi-person pose estimation is a fundamental task with many huge challenges in computer vision, such as close-interaction scenarios, limb occlusion. In this paper, we present two novel modules to group invisible keypoints occluded by various objects into the right body part. First, the Attention Residual Bottleneck is designed to generate high-resolution representations with enhanced channel-wise and spatial contextual information, which is integrated into the original residual unit with an attention mechanism. Second, the Object De-Occlusion Module is proposed to inference the order of object occlusion via occlusion relationship recovery strategy, predict and complete content for the invisible region of each person instance. Our proposed modules are evaluated on the COCO2017 keypoint benchmark, and experimental results show that our model has greater performance and faster inference speed compared to most of previous methods. It achieves the balance between accuracy and speed.