Multi-Person Pose Estimation using an Ordinal Depth-Guided Convolutional Neural Network

Monocular 2D multi-person pose estimation in videos is essential for applications such as surveillance, action recognition, kinematics analysis, and medical diagnosis. Existing state-of-the-art offsets-based methods extract temporal features from offsets in consecutive predicted rough skeletons for better preciseness in fine-tuned the skeletons. However, the precision of existing single image-based models of rough skeleton prediction, such as HRNet, dropped due to shifting of target persons in propagated bounding boxes and resulted in inconsistent estimated poses in consecutive frames. To conquer this problem, we proposed an Ordinal Depth-Guided-Convolutional Neural Network (ODG-CNN) to address the issue. The proposed ordinal depth guides the Ordinal Depth-Guided Block (ODGB) in the ODG-CNN to reweight features for target persons in bounding boxes. Experiment results on the PoseTrack 2018 dataset indicate that the proposed ODG-CNN achieves the highest performance in terms of mean Average Precision (mAP). The proposed ODG-CNN is suited for applications, such as use of telehealth for early detection and intervention of developmental delays in children, which needs high accuracy of video-based estimated poses.