INSID3 Beats Multi-Model Pipelines 30x Faster With One Frozen Backbone
INSID3 uses a single frozen DINOv3 backbone to segment any object from one example, with no training, no SAM, and no decoder — beating prior work by 7.5% mIoU
Takeaways
- INSID3 performs one-shot in-context segmentation using only a single frozen DINOv3 backbone — no SAM, no decoder, no training.
- It outperforms prior state-of-the-art by +7.5% mIoU while using 3x fewer parameters and running ~30x faster than Matcher.
- Key discovery: DINOv3 features contain a hidden positional bias that corrupts cross-image matching; INSID3 removes it via a training-free PCA projection.
- Works across object-level, part-level, and personalized segmentation, and generalizes to medical, aerial, and underwater domains from one annotated example.
- Accepted as a CVPR 2026 Oral; code is open-source under Apache 2.0, compatible with PyTorch 2.0+.
- The positional bias fix has implications beyond segmentation — any DINOv3-based retrieval or correspondence system may benefit.
Most segmentation systems today are built like assembly lines: one model finds the object, another generates the mask, a third refines the edges. INSID3, a CVPR 2026 Oral paper from researchers at Politecnico di Torino, TU Darmstadt, and TU Munich, tears that pipeline down to a single component. Show it one annotated image of a concept, and it will segment that concept in any new image , with no fine-tuning, no segmentation decoder, and no auxiliary models like SAM.
The Problem with How We Do Segmentation Today
In-context segmentation (ICS) is the task of segmenting an arbitrary concept given just one visual example , a single image with a mask drawn around the thing you care about. It could be a specific dog breed, a tumor in a medical scan, or the wing of an airplane in an aerial photo. The model has to figure out what you mean from that one example alone.
Current approaches fall into two camps, and both have real drawbacks:
- Fine-tuned models: These are trained on labeled data for specific categories. They perform well in-domain but fall apart when you show them something outside their training distribution.
- Multi-model pipelines: Systems like Matcher combine a frozen feature extractor (like DINOv2) with SAM for mask generation. They generalize better, but the architecture is complex, slow, and locked to a fixed level of segmentation granularity , you can segment whole objects or parts, but not both with the same system.
The question the INSID3 team asked is deceptively simple: does any of that complexity actually need to exist?
One Backbone to Rule Them All
DINOv3 is Meta's latest self-supervised vision backbone. It introduces Gram anchoring to solve dense feature degradation during long training schedules, scales up to 6.7B parameters, and trains on 1.7 billion web images plus 493 million satellite images. The result is a versatile vision foundation model that outperforms the specialized state of the art across a broad range of settings, without fine-tuning, producing high-quality dense features that achieve outstanding performance on various vision tasks.