Reproduce “Feature Puzzles: Single-Image Supervision for Multi-View 3D-Aware Segmentation” 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
Jiahao Ma, Qiang Zhang, Miaomiao Liu, David Ahmedt-Aristizabal, Chuong Nguyen
Feature Puzzles trains promptable multi-view segmentation from a single labeled image. It creates overlapping crops, transfers shared-parent labels between them, and uses a pretrained 3D-aware backbone to provide geometric features; it does not learn geometry from the image alone. On NVOS, it reaches 93.2% mIoU versus OmniSeg3D’s 92.8%, but on SPIn-NeRF its 88.5% is below leading methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.