Title: Geometry-Guided Modeling of Foundation Features Enables Generalizable Object Shape Deformation Learning

Abstract: While monocular 3D shape recovery is essential for geometric comprehension, ensuring robust generalization across diverse viewpoints and unseen object categories continues to pose a major challenge. This study introduces a deformable learning framework capable of reconstructing 3D objects by explicitly warping a category-level shape template to align with the observed target. To manage the intricate shape discrepancies between the template and the target, we propose a geometry-guided feature modeling mechanism. This approach initially enhances foundation features with template topology to create a geometry-aware representation, which is then explicitly correlated with the target observation to direct accurate deformation. Additionally, we address the gap between the static template and variable target perspectives through a view-adaptive feature aggregation module. By utilizing multi-view template features alongside their associated camera poses, this module enriches the canonical template representation, guaranteeing robust feature alignment irrespective of the target’s viewpoint. Comprehensive experiments reveal that our method surpasses existing state-of-the-art techniques in managing significant shape variations and diverse viewing angles, demonstrating strong generalization capabilities to novel categories and effectively facilitating downstream real-world dexterous robotic manipulation tasks.

Project homepage: https://GODeform.github.io/

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC