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

EOVSAM is an efficient open-vocabulary segmentation framework built on SAM 3 that adapts SAM 3 for single-pass prediction.

• Paper: EOVSAM: Efficient Open-Vocabulary Segmentation with SAM 3 in One Pass
• Code: GitHub Repository

Overview

EOVSAM removes prompt conditioning to turn SAM 3 into an efficient mask generator and introduces an Attentional Aggregation strategy to optimize open-vocabulary classification end-to-end. This formulation avoids multi-stage pipelines and post-processing heuristics while consistently improving segmentation accuracy over vanilla SAM 3 and accelerating inference by up to 338×.

Usage

Please refer to the official EOVSAM GitHub repository for installation, dataset preparation, training, and evaluation scripts.

License

This project is a composite distribution incorporating NVIDIA RADIO, SAM 3, and MAFT-Plus components; each remains subject to its upstream terms. Please check the GitHub License section for details.

Citation

@misc{peng2026eovsamefficientopenvocabularysegmentation,
      title={EOVSAM: Efficient Open-Vocabulary Segmentation with SAM 3 in One Pass}, 
      author={Haomin Peng and Yongkang Li and Zhaoxiang Liu and Xiaojie Jin and Shiguo Lian and Yunchao Wei and Xinggang Wang},
      year={2026},
      eprint={2608.02284},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.02284}, 
}