DIGS decouples instance discovery from linguistic naming.
Distribution-aware Gaussian labels enable training-free open-vocabulary 3D segmentation
The authors replace per-Gaussian single features with explicit top-M label distributions, recovering clean instance boundaries and cutting scene mapping time from hours to minutes.
Research Lab
Zhifan Zeng · Wenbo Xiao · Arcot Sowmya · Changming Sun
UNSW Sydney · CSIRO
Research Digest··3 min read
Zeng et al.
Why this paper
From CSIRO and UNSW Sydney
In one line
DIGS preserves label distributions per Gaussian and delays label selection, delivering faster training-free open-vocabulary 3D segmentation with cleaner boundaries and leading benchmark accuracy.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (hardware A100)
- ✓Limitations stated by the authors (2 noted)
- ✓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.
§