Semantic-Fast-SAM: Efficient Semantic Segmenter

July 8, 2026 ยท View on GitHub

Official implementation of Semantic-Fast-SAM: Efficient Semantic Segmenter.

Author: Byunghyun Kim
arXiv: https://arxiv.org/abs/2604.20169
arXiv DOI: https://doi.org/10.48550/arXiv.2604.20169
IEEE Xplore: https://ieeexplore.ieee.org/document/11249315
IEEE DOI: https://doi.org/10.1109/APSIPAASC65261.2025.11249315

Paper

Semantic-Fast-SAM: Efficient Semantic Segmenter
Byunghyun Kim

Semantic-Fast-SAM (SFS) is an efficient semantic segmentation framework that combines FastSAM mask generation with semantic labeling. It produces semantic segmentation maps with substantially lower computational cost than SAM-based semantic segmentation pipelines, while retaining the segment-anything capability for closed-set and open-vocabulary segmentation.

The paper is available as an arXiv preprint and as an IEEE APSIPA ASC proceedings paper.

  • Preprint: arXiv:2604.20169 [cs.CV]
  • Published version: Proceedings of the 2025 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), IEEE, Singapore, 2025, pp. 1158-1163

Installation

Clone this repository:

git clone https://github.com/KBH00/Semantic-Fast-SAM.git
cd Semantic-Fast-SAM

Create and activate the conda environment:

conda env create -f environment.yaml
conda activate sfs

Install the required spaCy model:

python -m spacy download en_core_web_sm

Download the FastSAM checkpoint from the following link:

https://drive.google.com/file/d/1l7l1VJmpD1nOsgiTXucTtYOpu3nE-rjh/view?usp=sharing

Place the checkpoint in the weights/ directory. If the downloaded file is named FastSAM-x.pt, rename it to FastSAM.pt.

Expected path:

weights/FastSAM.pt

Inference

Run inference with:

python scripts/main_ssa_engine.py \
  --data_dir data/<image_or_directory_name> \
  --out_dir output \
  --world_size <number_of_gpus>

For example:

python scripts/main_ssa_engine.py \
  --data_dir data/example.jpg \
  --out_dir output \
  --world_size 1

You can also run scripts/main_ssa_engine.py directly after setting the required arguments in the script.

Examples

Semantic-Fast-SAM example: cat

Semantic-Fast-SAM example: dogs

Citation

If you use this repository, please cite the paper.

arXiv version

@article{kim2026semanticfastsam,
  title        = {Semantic-Fast-SAM: Efficient Semantic Segmenter},
  author       = {Kim, Byunghyun},
  journal      = {arXiv preprint arXiv:2604.20169},
  year         = {2026},
  doi          = {10.48550/arXiv.2604.20169},
  url          = {https://arxiv.org/abs/2604.20169}
}

IEEE APSIPA ASC version

@inproceedings{kim2025semanticfastsam,
  title        = {Semantic-Fast-SAM: Efficient Semantic Segmenter},
  author       = {Kim, Byunghyun},
  booktitle    = {Proceedings of the 2025 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)},
  pages        = {1158--1163},
  year         = {2025},
  publisher    = {IEEE},
  address      = {Singapore},
  doi          = {10.1109/APSIPAASC65261.2025.11249315},
  url          = {https://doi.org/10.1109/APSIPAASC65261.2025.11249315}
}

Paper Metadata

For better discoverability and citation-tool indexing of the arXiv version, keep the paper metadata consistent across this repository, the arXiv page, the IEEE page, and any project website:

  • Title: Semantic-Fast-SAM: Efficient Semantic Segmenter
  • Author: Byunghyun Kim
  • arXiv ID: 2604.20169
  • arXiv DOI: 10.48550/arXiv.2604.20169
  • IEEE DOI: 10.1109/APSIPAASC65261.2025.11249315

A dedicated project page with Google Scholar citation meta tags is recommended for stronger indexing. For example, a GitHub Pages site can expose citation_title, citation_author, citation_publication_date, citation_doi, citation_arxiv_id, and citation_pdf_url fields in the HTML header.

This project builds on the following repositories:

License

Please check the license terms of this repository and the related projects before using the code, models, or checkpoints.