Reproduce “Revisiting Triplet Contrastive Learning for 3D Shape Understanding” 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
Chaoyang Zhu, Zhe Liao, Jiayi Zhang, Jie Li, Jiayi Ji, Liujuan Cao
HyperCL trains 3D shape representations using text rather than rendered-image features. It maps text and point-cloud embeddings into hyperbolic space and uses contrastive and entailment losses to place specific shapes under general descriptions. On c-OLVIS it reaches 86.2% Acc@1, versus 81.6% for Uni3D-B. In text-to-3D retrieval without LVIS, it scores 25.2% Recall@1 versus TAMM’s 21.9%, in that test setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.