Research map / Detection and segmentation / Weakly supervised detection and segmentation
INSID3: Training-Free In-Context Segmentation with DINOv3
Claudia Cuttano, Gabriele Trivigno, Christoph Reich, Daniel Cremers, Carlo Masone, Stefan Roth
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use frozen DINOv3 features only (no SAM, no fine-tuning) for in-context segmentation | INSID3 · One-shot ICS average (9 datasets) | mIoU: 47.6 -> 55.1 vs GF-SAM | Improved |
| swap DINOv2 for DINOv3 in SAM-based GF-SAM (ablation) | GF-SAM · One-shot ICS average (9 datasets) | mIoU: 47.6 -> 47.0 | No clear effect |
| add positional debiasing to GF-SAM with DINOv3 (ablation) | GF-SAM (DINOv3 + SAM) · One-shot ICS average (9 datasets) | mIoU: 47.0 -> 48.8 | Improved |
| add positional debiasing of DINOv3 features for cross-image matching (ablation) | DINOv3 on SPair-71k · SPair-71k PCK (Base, T=0.20) | PCK: 59.8 -> 66.4 | Improved |
| add positional debiasing with rank s=500 (ablation) | INSID3 · COCO-20i and PASCAL-Part | mIoU: +3.1 % on COCO, +2.7 % on PASCAL | Improved |
| remove too many dimensions in the debiasing subspace (large rank s) (ablation) | INSID3 · COCO-20i and PASCAL-Part | mIoU | Hurt |
| use finer clustering (larger tau) (ablation) | INSID3 · Part vs object segmentation | mIoU | Mixed |
| add agglomerative clustering instead of thresholding the similarity map (ablation) | INSID3 · COCO-20i | mIoU: 44.2 -> 50.6 (tau=0.5, single cluster) | Improved |
| add cluster aggregation with cross-image similarity on top of fine clustering (ablation) | INSID3 · PASCAL-Part | mIoU: 36.2 -> 48.5 | Improved |
| add self-similarity to cross-similarity in aggregation score (ablation) | INSID3 · COCO-20i | mIoU: 54.6 -> 57.6 | Improved |
| add self-similarity to cross-similarity in aggregation score (ablation) | INSID3 · PASCAL-Part | mIoU: 48.5 -> 50.5 | Improved |
| fine-tune on COCO (SegIC) for in-domain gains | SegIC · COCO-20i vs out-of-domain | mIoU: COCO-20i 76.1 vs ISIC 25.3 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- SegIC — SEGIC: Unleashing the Emergent Correspondence for In-Context Segmentation (ECCV 2024) · better on 2 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SegGPT: Towards Segmenting Everything in Context (ICCV 2023 · similarity 0.87)
- SANSA: Unleashing the Hidden Semantics in SAM2 for Few-Shot Segmentation (NeurIPS 2025 · similarity 0.86)
- UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · similarity 0.86)
- Training-free Open-Vocabulary Semantic Segmentation via Diverse Prototype Construction and Sub-region Matching (AAAI 2025 · similarity 0.86)
- Semantic-aware SAM for Point-Prompted Instance Segmentation (CVPR 2024 · similarity 0.86)
- Explore In-Context Segmentation via Latent Diffusion Models (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
