Reproduce “VINO: Revisiting 3D Representation Learning with Volume Transformers” 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
Adrian Kruse, Kadir Yilmaz, Daan de Geus, Francis Engelmann, Bastian Leibe
VINO learns 3D scene features without human-labeled pretraining data, using teacher views, corrupted student views, masked-token prediction from visible context, and image targets consistent across camera views. On ScanNet200 linear probing, its 3D-only first stage reaches 35.9 mIoU, 5.2 points above DOS. The probe uses labeled downstream data; the final model also uses posed RGB images and a frozen DINOv2 teacher.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.