Reproduce “Affostruction: 3D Affordance Grounding with Generative Reconstruction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chunghyun Park, Seunghyeon Lee, Minsu Cho
Affostruction is a generative framework that reconstructs complete 3D object geometry from partial multi-view RGBD observations and grounds open-vocabulary affordances on the full shape. It achieves a 67.6% improvement in volumetric IoU for 3D reconstruction over TRELLIS and nearly doubles the aIoU for partial 3D affordance grounding compared to leading baselines.
Key insights
- Integrating depth conditioning and sparse voxel fusion into flow matching-based generation significantly enhances 3D reconstruction and extrapolation of complete geometry from partial views.
- Generative modeling of affordances, instead of discriminative prediction, enables a more comprehensive understanding by capturing multiple valid interaction regions.
- Leveraging predicted affordances to actively guide view selection allows for more efficient data acquisition, accelerating the process of gaining a full functional understanding of objects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.