DINOde: Continuous Vision-Text Alignment for Open-Vocabulary Semantic Segmentation

Researchers introduced DINOde, an ODE-based framework for continuous vision-text alignment in open-vocabulary semantic segmentation.

It maps CLIP text embeddings and global image features into DINO’s visual space through Semantic Text Flow and Global Context Flow.

The method also uses tangent-space constraints to preserve hyperspherical geometry and improve cross-modal consistency.

DINOde reported state-of-the-art results on OVSS benchmarks, strengthening recognition of unseen categories.