Home

Awesome

Surgical SAM 2: Real-time Segment Anything in Surgical Video by Efficient Frame Pruning

Official implementation for SurgSAM2, an innovative model that leverages the power of the Segment Anything Model 2 (SAM2), integrating it with an efficient frame pruning mechanism for real-time surgical video segmentation. Paper in arxiv.

We introduce Surgical SAM 2 (SurgSAM-2), an innovative model that leverages the power of the Segment Anything Model 2 (#SAM2), integrating it with an efficient frame pruning mechanism for real-time surgical video segmentation. The proposed SurgSAM-2

architecture

Training

The source code is coming soon

Evaluation data and pretrained weighted

The demo data from Endovis 2018 could be downloaded from 2018 demo data. Please put the data to construct the data construction as: project_root/datasets/endovis18/images/seq_2/...

The pretrained weighted could be downloaded from sam2_hiera_s_endo18. Please put the weight to construct project_root/model_weights/sam2_hiera_s_endo18.pth

Acknowledgement

This model was trained with the datasets from Endovis 2017, Endovis 2018. If you need to use the data, please apply for the usage from their website Endovis 2017 and Endovis 2018.

This code was adapted from segment anything 2. We are grateful for their excellent code and contribution to video segmentation.

Citation

@misc{liu2024surgicalsam2realtime,
 title={Surgical SAM 2: Real-time Segment Anything in Surgical Video by Efficient Frame Pruning}, 
author={Haofeng Liu and Erli Zhang and Junde Wu and Mingxuan Hong and Yueming Jin},
 year={2024},
 eprint={2408.07931},
 archivePrefix={arXiv},
 primaryClass={cs.CV},
 url={https://arxiv.org/abs/2408.07931}, 
}