Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a training-free, two-stage framework for fine-grained semantic segmentation in low-data regimes, specifically addressing the FungiTastic dataset's challenges of long-tailed distributions and variable image conditions. The approach leverages SAM for class-agnostic mask generation using macro-taxonomic prompts, followed by DINOv3 for fine-grained label assignment via prototype matching. A simple transformation of the DINOv3 feature space further enhances prototype-based classification accuracy, establishing a strong baseline for few-shot semantic segmentation.
Achieve surprisingly strong few-shot semantic segmentation by simply combining SAM's masking with DINOv3's feature space and a clever transformation, no training required.
Fine-grained semantic segmentation requires both precise localization and discrimination between visually similar classes. In FungiTastic, this problem is further complicated by a long-tailed distribution and strong variation in image acquisition conditions. We propose a training-free two-stage framework that decouples segmentation from classification. SAM3 first produces class-agnostic mushroom masks using macro-taxonomic prompts, and DINOv3 then assigns fine-grained labels through prototype matching in the embedding space. To improve this stage, we apply a simple transformation of the DINOv3 feature space that improves prototype-based classification. Compared with class-specific prompting, our approach is more scalable and keeps the segmentation cost low. We report results from one-shot to few-hundred-shot regimes, providing, to the best of our knowledge, the first baseline for fine-grained semantic segmentation in low-data settings.