Reproduce “UniT: Unified Geometry Learning with Group Autoregressive Transformer” 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
Haotian Wang, Yusong Huang, Zhaonian Kuang, Hongliang Lu, Xinhu Zheng, Meng Yang, Gang Hua
Researchers at HKUST and Xi'an Jiaotong University developed UNIT, a unified feed-forward model capable of integrating diverse geometry perception functionalities, including online, offline, multi-modal, long-horizon, and metric-scale tasks. The model achieved state-of-the-art performance across ten benchmarks and seven tasks while maintaining bounded memory usage for extended sequences.
Key insights
- A single Group Autoregressive Transformer can unify diverse geometry perception capabilities by dynamically adjusting the group size for varying view configurations.
- Bounded memory for long-horizon perception is achievable via anchor-free relational modeling and queue-style KV caching, while multi-modal fusion benefits from spatially aligned concatenation in modal attention.
- Generalization in metric-scale learning can be improved through a curriculum-like loss design that initially focuses on scale-invariant geometry before progressively guiding the model to learn absolute metric scale.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.