Reproduce “MVT: Multi-view Vision Transformer for 3D Object Recognition” 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
Shuo Chen, Tan Yu, Ping Li
MVT lets patches from rendered views interact through ordinary Transformer attention before pooling. It processes patches within each view first, then concatenates view features for global Transformer blocks, avoiding one all-patch sequence. On six-view ModelNet10, the tiny scratch-trained model reaches 92.35% with four global blocks, versus 90.42% with none; this supports, but does not prove, that cross-view communication helps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.