Overview
Segmenting a new biomedical dataset typically demands either a domain-specific model trained on extensive annotations or a foundation model guided at inference time. We introduce Exemplar, a few-shot segmentation method that combines a frozen DINOv3 backbone with a fixed bank of classical, native-resolution filter responses within a single lightweight head, fitted solely from support masks. In the few-mask, native-resolution regime, classical priors and frozen self-supervised features prove complementary: fused in one head, a single fixed configuration generalizes across eleven biomedical imaging datasets.
Key Contributions
- Unified architecture: Exemplar merges classical and learned features into one head, eliminating the need for fine-tuning the backbone.
- State-of-the-art few-shot performance: Outperforms five forward-pass methods in 54 of 55 method-dataset comparisons, with 52 remaining significant after Holm correction.
- Data efficiency: From a single annotated mask, Exemplar achieves 0.703 on a benchmark panel, surpassing a from-scratch nnU-Net (0.682) trained on the same mask. At eight masks, nnU-Net overtakes the panel mean—chiefly on centreline agreement—but requires 16–77× longer fitting time.
- Complementarity of priors: Under the same head, the classical bank alone reaches 0.693 (foreground intersection-over-union or centreline Dice), while the frozen features alone score 0.672. The classical bank leads on seven datasets, features on four, and their fusion reaches 0.782.
Method in Brief
Exemplar operates at native resolution, preserving fine structural details critical for microscopy. It freezes a DINOv3 feature extractor and a handcrafted filter bank (e.g., edge, texture, and morphological operators). The lightweight head learns a mapping from concatenated features to segmentation masks using only the support set, without backpropagating through the feature extractors.
Results and Impact
Exemplar demonstrates that classical image-processing priors—often overlooked in favor of end-to-end learning—meaningfully complement modern self-supervised representations in data-scarce settings. This opens practical doors for biomedical researchers needing rapid, accurate segmentation with minimal manual annotation.
The code, configurations, and per-image score records are available at https://github.com/michalprusek/Exemplar.
via ArXiv CV