Open-vocabulary semantic segmentation enables text-driven segmentation of novel classes, but fine-grained remote sensing scenarios remain challenging due to visual similarity, semantic ambiguity, and blurry object boundaries. SANO3 is a training-free framework that enhances SAM 3 with refined textual and spatial guidance. It leverages the visual self-similarity priors of vision foundation models to extract discriminative visual primitives, aligns them with text queries through visual-conditioned relation-aware alignment, and injects hybrid visual prototypes for deeper visual-linguistic fusion. To further refine boundaries, SANO3 formulates spatial prompt generation as an iterative information-gain process that greedily selects informative point prompts via entropy reduction.
We introduce FG-OVSSRS, a fine-grained open-vocabulary semantic segmentation benchmark for remote sensing images. The benchmark covers diverse remote-sensing scenarios and provides synonym-expanded category queries to evaluate model robustness under flexible real-world open-vocabulary usage.
The benchmark is publicly available on Hugging Face: 🤗 FG-OVSSRS Bench
- Release the FG-OVSSRS benchmark.
- Release SANO3 code.