EOVSAM: Efficient Open-Vocabulary Segmentation with SAM 3 in One Pass

Researchers introduced EOVSAM, a framework that repurposes SAM 3 for open-vocabulary segmentation without prompt conditioning.

By adding attentional aggregation, EOVSAM enables efficient single-pass prediction instead of slow multi-stage processing.

The method outperforms vanilla SAM 3 on evaluated semantic and panoptic segmentation benchmarks.

Inference is reported to be up to 338x faster, making large-scale text-guided segmentation more practical.