Mask-Adapter

The Devil is in the Masks for Open-Vocabulary Segmentation

Yongkang Li^1,*, Tianheng Cheng^1,*, Wenyu Liu¹, Xinggang Wang^1,📧

¹ Huazhong University of Science and Technology,

(* equal contribution, 📧 corresponding author)

Highlights

• Mask-Adapter is a simple yet remarkably effective method and can be seamlessly integrated into open-vocabulary segmentation methods, e.g., FC-CLIP and MAFT-Plus, to tackle the existing bottlenecks.

• Mask-Adapter effectively extends to SAM without training, achieving impressive results across multiple open-vocabulary segmentation benchmarks.

Updates

• Release code
• Release weights
• Release demo with SAM-2👉 🤗 Mask-Adapter
• Release weights training with addtional data

Getting Started

• Installation.

• Preparing Datasets for Mask-Adapter.

• Getting Started with Mask-Adapter.

Models

Model	Backbone	A-847	A-150	PC-459	PC-59	PAS-20	Download
FC-CLIP	ConvNeXt-L	14.8	34.1	18.2	58.4	95.4	model
FC-CLIP + Mask-Adapter	ConvNeXt-L	14.1	36.6	19.3	59.7	95.5	model
MAFTP-Base	ConvNeXt-B	13.8	34.5	18.5	57.5	95.5	model
MAFTP-Base + Mask-Adapter	ConvNeXt-B	14.2	35.6	17.9	58.4	95.1	model
MAFTP-Large	ConvNeXt-L	15.5	36.3	21.2	59.5	96.4	model
MAFTP-Large + Mask-Adapter	ConvNeXt-L	16.2	38.2	22.7	60.4	95.8	model

Citation

If you Mask-Adapter useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.

@article{li2024maskadapter,
      title={Mask-Adapter: The Devil is in the Masks for Open-Vocabulary Segmentation}, 
      author={Yongkang Li and Tianheng Cheng and Wenyu Liu and Xinggang Wang},
      year={2024},
      eprint={2412.04533},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2412.04533}, 
}

License

All code in this repository is under the Apache License 2.0.

Acknowledgement

Mask-Adapter is based on the following projects: detectron2, Mask2Former, FC-CLIP and MAFTP. Many thanks for their excellent contributions to the community.