REBASE: Reference-Background Subspace Elimination for Training-Free In-Context Segmentation

Mantha Sai Gopal, Jaison Saji Chacko, Harsh Nandwana, Sandesh Hegde, Debarshi Banerjee, Uma Mahesh

CamCom Technologies Private Limited

Training-free in-context segmentation. A single reference image and its binary mask are sufficient to segment the same category in any query image—no gradient updates or per-target adaptation required.

Method

Results

Method	ISIC	X-Ray	FSS-1000	PASCAL-Part	PACO-Part
PerSAM	23.9	31.7	71.2	32.5	22.5
Matcher	38.6	70.8	87.0	42.9	34.7
GF-SAM	48.7	51.0	88.0	44.5	36.3
INSID3 (DINOv3-L)	54.4	78.8	83.7	50.5	38.7
REBASE (ours)	63.6	87.8	88.6	46.1	39.6

1-shot mIoU (%, ↑). All training-free methods except INSID3 use DINOv2-L.

Code

Code release coming soon. ⭐ star the repo for updates.

Citation

@article{gopal2026rebase,
  title   = {REBASE: Reference-Background Subspace Elimination for Training-Free In-Context Segmentation},
  author  = {Mantha Sai Gopal and Jaison Saji Chacko and Harsh Nandwana and Sandesh Hegde and Debarshi Banerjee and Uma Mahesh},
  journal = {arXiv preprint arXiv:2607.09082},
  year    = {2026}
}

Acknowledgements

This project builds upon Segment Anything (SAM) and DINOv2. We also thank the authors of PerSAM, Matcher, GF-SAM, and INSID3 for their open-source implementations and benchmarks.