Sub-Semantic Image Segmentation

The paper introduces a novel category of sub-semantic image segmentation, which uses language to partition images into stable appearance patterns rather than naming whole objects. The authors present DETECTURE, a method that addresses issues such as language leakage and prompt competition when coupling a vision-language model with the SAM 3 segmentation backbone. Additionally, they introduce the TextureADE dataset for evaluating this approach, demonstrating that DETECTURE outperforms existing baselines across multiple metrics. This advancement is significant for practitioners as it enhances the granularity of image segmentation tasks, allowing for more nuanced visual understanding in AI applications.