INSID3: Training-Free In-Context Segmentation Matches SAM Pipelines Using Only Frozen DINOv3

INSID3 is a training-free in-context segmentation method that segments objects, parts, and personalized instances from a single annotated reference image using only frozen DINOv3 features — no segmentation decoder, no fine-tuning, and no auxiliary models like SAM. The approach works by combining agglomerative clustering on dense patch embeddings, a positional debiasing step via SVD projection to suppress spurious cross-image coordinate matches, and a seed-cluster aggregation stage that recovers full object extent from a single reference mask.

Across nine benchmarks, INSID3 achieves 55.1 mIoU, outperforming SAM-based baselines by 8.1 points while running 3.4× faster and using 3× fewer parameters (304M vs. ~945M). The positional debiasing technique also generalizes beyond segmentation: on the SPair-71k semantic correspondence benchmark, it lifts PCK@0.10 by 5.8 points over vanilla DINOv3 features, suggesting broader utility for dense visual matching tasks.