Existing triplet contrastive pretraining aligns 3D shapes with both language and rendered 2D views, but the authors argue that rendered images carry distorted semantics due to domain gap, self-occlusion and rendering artifacts.
Language-only hyperbolic contrastive learning improves 3D shape understanding over image-based triplets
HyperCL aligns 3D shapes with text in hyperbolic space and enforces a partial order, outperforming triplet methods that rely on rendered images.
Chinese Tech
Chaoyang Zhu · Zhe Liao · Jiayi Zhang · Jie Li · Jiayi Ji · Liujuan Cao
Xiamen University · Shanghai AI Lab
Research Digest··2 min read
Zhu et al.
Why this paper
From Shanghai AI Lab and Xiamen University
In one line
Learning 3D representations from language alone in hyperbolic space beats triplet alignment with rendered images on classification and retrieval.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§