Event

Peking University's DragMesh Accepted to IJCV 2026

1 reports · 1 sources

DragMesh: Interactive 3D Generation Made Easy, proposed by a Peking University team, has been accepted to the International Journal of Computer Vision (IJCV). The method targets local interaction with existing 3D assets: through a single drag, it recovers the joint type, motion axis, and joint origin of the target part, and generates a complete and continuous joint motion trajectory, without the need to retrain a motion model for each new object. The paper focuses primarily on single-joint rigid articulated objects and validates under the clean mesh setting of GAPartNet. Its kinematic corpus contains 23 categories and 1045 articulated objects, of which 941 are used for KPP-Net training and 104 for object-disjoint validation. The 85 FPS mentioned in the paper refers only to the speed of the local DQ-VAE decoding, semantic skinning, and rendering stages after static 3DGS reconstruction is complete, and does not include segmentation preprocessing or optional VLM API latency.