Peking University's DragMesh turns static 3D assets into articulated objects with one drag, accepted to IJCV
jiqizhixin · x · 2026-10-31
Peking University's DragMesh converts a static 3D asset into an articulated object from a single drag, with no per-object motion model training; the work is reported accepted to IJCV.
- Infer kinematics first: semantic intent plus geometric prediction yields joint type, axis and origin (drawers slide, doors rotate) — the drag alone doesn't specify these constraints.
- Generate coherent trajectories: a Dual Quaternion VAE jointly represents rotation and translation, with FiLM conditioning repeatedly injecting joint information into a non-autoregressive Transformer decoder so motion respects the inferred kinematics.
- Pipeline-friendly: works alongside text-to-3D models like Hunyuan3D 2.5/TripoSG/Rodin plus MeshAnythingV2, part segmentation, and a KPP network that predicts joint axis and origin.
Experiments validate articulation on curated assets, targeting interactive 3D assets and simulation.
More from Research
- LLM judges harbor systematic errors that silently corrupt model rankings and training pipelines — burkov · 2026-10-31
- New paper quantifies ineffable concepts: 65 of 1,270 can never be taught to LLMs via language — sanmikoyejo · 2026-10-31
- Unsupervised Method Measures Whether LLM Messages Are Received — sanmikoyejo · 2026-10-31
- Neural net flies a drone in Acro mode from streamed game footage on a tiny board — yacineMTB · 2026-10-31
- llmlip-rs: Rust MLIP inference engine runs 6000-atom sims 17.7x faster than PyTorch MACE — CatAstro_Piyush · 2026-10-31
- AdaPRA: a process-reward agent with adaptive rubrics audits data science agents as they work — dan_fried · 2026-10-31