CVPR 2026 FGVC Workshop

Training-Free Fine-Grained Semantic Segmentations in Low Data Regimes: A FungiTastic Baseline

Sebastian Cavada, Francesco Pelosin, Lapo Faggi

Abstract

Fine-grained semantic segmentation in FungiTastic requires accurate mask localization while distinguishing between visually similar mushroom categories under long-tailed data and variable capture conditions. This work introduces a training-free two-stage baseline: SAM3 first generates class-agnostic mushroom masks from macro-taxonomic prompts, then DINOv3 assigns fine-grained labels using prototype matching in feature space. A simple feature-space transformation improves the prototype classifier, making the approach more scalable than class-specific prompting while keeping segmentation cost low. The paper reports performance from one-shot to few-hundred-shot settings and establishes an initial baseline for low-data fine-grained semantic segmentation.