Skip to content

Latest commit

 

History

History
282 lines (256 loc) · 8.38 KB

README.md

File metadata and controls

282 lines (256 loc) · 8.38 KB
Raw

RepViT-SAM: Towards Real-Time Segmenting Anything

Official PyTorch implementation of RepViT-SAM, from the following paper:

RepViT-SAM: Towards Real-Time Segmenting Anything.
Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding
[arXiv]


Models are deployed on iPhone 12 with Core ML Tools to get latency.

Abstract Segment Anything Model (SAM) has shown impressive zero-shot transfer performance for various computer vision tasks recently. However, its heavy computation costs remain daunting for practical applications. MobileSAM proposes to replace the heavyweight image encoder in SAM with TinyViT by employing distillation, which results in a significant reduction in computational requirements. However, its deployment on resource-constrained mobile devices still encounters challenges due to the substantial memory and computational overhead caused by self-attention mechanisms. Recently, RepViT achieves the state-of-the-art performance and latency trade-off on mobile devices by incorporating efficient architectural designs of ViTs into CNNs. Here, to achieve real-time segmenting anything on mobile devices, following, we replace the heavyweight image encoder in SAM with RepViT model, ending up with the RepViT-SAM model. Extensive experiments show that RepViT-SAM can enjoy significantly better zero-shot transfer capability than MobileSAM, along with nearly $10\times$ faster inference speed.

Installation

pip install -e .
# download pretrained checkpoint
mkdir weights && cd weights
wget https://github.com/THU-MIG/RepViT/releases/download/v1.0/repvit_sam.pt

Demo

Our Hugging Face demo is here

python app/app.py

CoreML export

Please refer to coreml_example.ipynb

Latency comparisons

Comparison between RepViT-SAM and others in terms of latency. The latency (ms) is measured with the standard resolution of 1024 $\times$ 1024 on iPhone 12 and Macbook M1 Pro by Core ML Tools. OOM means out of memory.

Platform Image encoder Mask decoder
RepViT-SAM MobileSAM ViT-B-SAM
iPhone 48.9ms OOM OOM 11.6ms
Macbook 44.8ms 482.2ms 6249.5ms 11.8ms

Zero-shot edge detection

Comparison results on BSDS500.

Model zero-shot edge detection
ODS OIS AP
ViT-H-SAM .768 .786 .794
ViT-B-SAM .743 .764 .726
MobileSAM .756 .768 .746
RepViT-SAM .764 .786 .773

Zero-shot instance segmentation and SegInW

Comparison results on COCO and SegInW.

Model zero-shot instance segmentation SegInW
AP $AP^{S}$ $AP^{M}$ $AP^{L}$ Mean AP
ViT-H-SAM 46.8 31.8 51.0 63.6 48.7
ViT-B-SAM 42.5 29.8 47.0 56.8 44.8
MobileSAM 42.7 27.0 46.5 61.1 43.9
RepViT-SAM 44.4 29.1 48.6 61.4 46.1

Zero-shot video object/instance segmentation

Comparison results on DAVIS 2017 and UVO.

Model z.s. VOS z.s. VIS
$\mathcal{J\&F}$ $\mathcal{J}$ $\mathcal{F}$ AR100
ViT-H-SAM 77.4 74.6 80.2 28.8
ViT-B-SAM 71.3 68.5 74.1 19.1
MobileSAM 71.1 68.6 73.6 22.7
RepViT-SAM 73.5 71.0 76.1 25.3

Zero-shot salient object segmentation

Comparison results on DUTS.

Zero-shot anomaly detection

Comparison results on MVTec.

Model z.s. s.o.s. z.s. a.d.
$\mathcal{M}$ $\downarrow$ $\mathcal{F}_{p}$
ViT-H-SAM 0.046 37.65
ViT-B-SAM 0.121 36.62
MobileSAM 0.147 36.44
RepViT-SAM 0.066 37.96

Acknowledgement

The code base is partly built with SAM and MobileSAM.

Thanks for the great implementations!

Citation

If our code or models help your work, please cite our paper:

@misc{wang2023repvitsam,
      title={RepViT-SAM: Towards Real-Time Segmenting Anything}, 
      author={Ao Wang and Hui Chen and Zijia Lin and Jungong Han and Guiguang Ding},
      year={2023},
      eprint={2312.05760},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}