ScholarWorks@숙명여자대학교

초록

Recent advancements in deep learning have improved quantitative accuracy in 3D human pose estimation, but the estimated poses occasionally suffer from visual defects such as joint tremors and protrusions. While existing 3D pose similarity metrics and estimation models managed to reduce visual defects by addressing the structure of human poses, they still struggle in scenarios where visually sensitive joints are prevalent, particularly in cases of self-occlusion. In this paper, we identify these visually sensitive joints and demonstrate the significance of explicitly considering structure and sensitivity in the problem of 3D human pose estimation. Building upon the successful consideration of human pose structure, we first propose a new enhanced pose similarity metric PSIM+, which models sensitivity similarity to further capture human perception and focus on visual defects. Furthermore, we introduce a new 3D pose estimation model Dual Graph-based Convolutional Neural Networks (DG-CNN), which reconstructs 3D poses by focusing on the spatio-temporal correlation of the skeletal structure and actively controlling visually sensitive joints. By incorporating a novel similarity loss function, our model can implicitly model the structure and sensitivity of human poses through its architecture and explicitly through direct supervision. Our model not only improves the accuracy of the estimated pose but also increases the perceptual quality as evaluated by PSIM+, verifying the significance of structure and sensitivity awareness. Through rigorous benchmarking, we demonstrate that our metric and estimation model achieve the highest correlation with user scores and perform best in situations where visually sensitive joints are prevalent.

키워드