曾爱玲.pptx

1、Human-centric Perception and Generation,Ailing ZengAnuttacon,【From video to motion】1)Markerless Motion capture algorithms for expressively,dynamically,and physically plausible 3D humans,2)Covering 2D/3D whole-body motions,3)Real-time motion capture process,pre-and post-processing smoothing,filtering

2、,denoising【From video/motion to language】Fine-grained motion and video understanding【Paired video/motion/text/audio】Obtain large-scale data from Internet data with the scalable and automatic capture system【From X to Motion,2D/3D/4D】Human-centric image/animation/motion/avatar generation【Human-in-the-

3、loop】Interactive techniques between AI and humans,The System for Large Motion Model with Scalable Data,Perception Understanding Data Generation Interaction,From Motion-X to Motion-X+for Large-scale Multimodal Human Motion Dataset,From SMPLer-X to SMPLest-X for 3D Expressive Human Pose and Shape Esti

4、mation,From DreamWaltz to DreamWaltz-G for 3D Gaussian Avatars,IDOL:Instant Photorealistic 3D Human Creation from a Single ImageCVPR 2025,SkillMimic:Learning Basketball Interaction Skills from Demonstrations,a,CVPR 2025 Highlight,Skill in Simulator,HOI Motion Data,Real-World Skill,MotionCapture,Skil

5、lMimic,Unified learning for all basketball skills,Purely data-driven,No case-by-case rewards,Overview,MoCap,Learning from Human-Object Interaction(HOI)Demonstrations,Skill Policy,Body Perception,Object Perception,Skill Condition,Motor Torques,Unified HOI Imitation Reward,Contact,Contact Graph,Unifie

6、d HOI Imitation Reward,Unified HOI Imitation Reward,Reinforcement Learning,Real World Skills,.,.,HOI Motion Dataset,Unified HOI Imitation Reward,.,.,Unified HOI Imitation Reward,Result,Unified HOI Imitation Reward,Dribble Left,Jump Shoot,Tu