mobilesam

Introduction

Official Repo

Code Snippet

MobileSAM (ArXiv'2023)

@article{mobile_sam,
    title={Faster Segment Anything: Towards Lightweight SAM for Mobile Applications},
    author={Zhang, Chaoning and Han, Dongshen and Qiao, Yu and Kim, Jung Uk and Bae, Sung-Ho and Lee, Seungkyu and Hong, Choong Seon},
    journal={arXiv preprint arXiv:2306.14289},
    year={2023}
}

Inference with MobileSAM

The usage of MobileSAM in sssegmenation is exactly the same as SAM by replacing

• SAM: MobileSAM,
• SAMPredictor: MobileSAMPredictor,
• SAMAutomaticMaskGenerator: MobileSAMAutomaticMaskGenerator.

Specifically, you can import the three classes by

from ssseg.modules.models.segmentors.mobilesam import MobileSAM
from ssseg.modules.models.segmentors.mobilesam import MobileSAMPredictor
from ssseg.modules.models.segmentors.mobilesam import MobileSAMAutomaticMaskGenerator

# predictor only could be MobileSAMPredictor(use_default_sam_t_5m=True, device='cuda')
predictor = MobileSAMPredictor(use_default_sam_t_5m=True, device='cuda')

# mask_generator only could be MobileSAMAutomaticMaskGenerator(use_default_sam_t_5m=True, device='cuda')
mask_generator = MobileSAMAutomaticMaskGenerator(use_default_sam_t_5m=True, device='cuda')

By the way, you can refer to inference-with-sam to learn about how to use SAM with sssegmenation. Also, you can refer to MobileSAM Official Repo to compare our implemented MobileSAM with official version.