• 강사 : 김태균 교수 (KAIST 전산학부)
• 일시 : 2025. 5. 26. (월) 16:00~17:30

This talk addresses denoising diffusion models for 3D human and object pose estimation under interactions. Diffusion models as a strong generative prior are integrated to pose estimation tasks from single image input. Multiple hypotheses conditioning helps accurate estimation under interactive occlusions, while diffusion models inpaint missing pixels in 3D (NeurIPS2024). Whole body and hand tracking is achieved by the cascaded diffusion model from stereo pairs of fisheye images from a head-mounted device. Despite significant occlusions of body parts in the views, the proposed method offers precise pose and 3D shape estimates of body and two hands. Casual temporal attention allows to extend to arbitrary long motion sequences (CVPR2025). Existing diffusion models for 6D object pose estimation require long learning time and an extra scoring network. We propose pre-training the encoder by direct pose regression and joint-learning by the pose regression head and diffusion head, dramatically reducing the learning time. The proposed score-scaling as diffusion guidance helps remove the scoring network, achieving on-par or beyond accuracies (ICCV2025 under review).