ZegCLIP: Towards Adapting CLIP for Zero-shot Semantic Segmentation

Authors: Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu*, Yifan Liu. *Corresponding author

[paper] [github] [docker image] [pretrained models] [visualization] [visualization of class queries]

---

Abstract: Recently, CLIP has been applied to pixel-level zero-shot learning tasks via a two-stage scheme. The general idea is to first generate class-agnostic region proposals and then feed the cropped proposal regions to CLIP to utilize its image-level zero-shot classification capability. While effective, such a scheme requires two image encoders, one for proposal generation and one for CLIP, leading to a complicated pipeline and high computational cost. In this work, we pursue a simpler-and-efficient one-stage solution that directly extends CLIP's zero-shot prediction capability from image to pixel level. Our investigation starts with a straightforward extension as our baseline that generates semantic masks by comparing the similarity between text and patch embeddings extracted from CLIP. However, such a paradigm could heavily overfit the seen classes and fail to generalize to unseen classes. To handle this issue, we propose three simple-but-effective designs and figure out that they can significantly retain the inherent zero-shot capacity of CLIP and improve pixel-level generalization ability. Incorporating those modifications leads to an efficient zero-shot semantic segmentation system called ZegCLIP. Through extensive experiments on three public benchmarks, ZegCLIP demonstrates superior performance, outperforming the state-of-the-art methods by a large margin under both ''inductive'' and ''transductive'' zero-shot settings. In addition, compared with the two-stage method, our one-stage ZegCLIP achieves a speedup of about 5 times faster during inference.

<p align="center"> <img width="800" src="figs/overview.png"> </p>

Environment:

Option 1:

• Install pytorch

conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio=0.10.1 cudatoolkit=10.2 -c pytorch

• Install the mmsegmentation library and some required packages.

pip install mmcv-full==1.4.4 mmsegmentation==0.24.0 pip install scipy timm==0.3.2

Option 2:

• Directly apply the same Image we provieded in Dockerhub:

docker push ziqinzhou/zegclip:latest

Downloading and preprocessing Dataset:

According to MMseg: https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/dataset_prepare.md

Preparing Pretrained CLIP model:

Download the pretrained model here: Path/to/ViT-B-16.pt https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt

Pretrained models:

| Dataset | Setting | pAcc | mIoU(S) | mIoU(U) | hIoU | Model Zoo | | :-------------: | :---------: | :---: | :-----: | :-----: | :--: | :----------------------------------------------------------: | | PASCAL VOC 2012 | Inductive | 94.6 | 91.9 | 77.8 | 84.3 | [Google Drive] | | PASCAL VOC 2012 | Transductive | 96.2 | 92.3 | 89.9 | 91.1 | [Google Drive] | | PASCAL VOC 2012 | Fully | 96.3 | 92.4 | 90.9 | 91.6 | [Google Drive] | | COCO Stuff 164K | Inductive | 62.0 | 40.2 | 41.1 | 40.8 | [Google Drive]| | COCO Stuff 164K | Transductive | 69.2 | 40.7 | 59.9 | 48.5 | [Google Drive]| | COCO Stuff 164K | Fully | 69.9 | 40.7 | 63.2 | 49.6 | [Google Drive] |

Note that here we report the averaged results of several training models and provide one of them.

Efficiency results:

| Dataset | #Params(M) | Flops(G) | FPS | | :-------------: | :---------: | :--------: | :--------: | | PASCAL VOC 2012 | 13.8 | 110.4 | 9.0 | | COCO Stuff 164K | 14.6 | 123.9 | 6.7 |

Note that all experience are conducted on a single 1080Ti GPU and #Params(M) represents the number of learnable parameters.

Training (Inductive):

bash dist_train.sh configs/coco/vpt_seg_zero_vit-b_512x512_80k_12_100_multi.py Path/to/coco/zero_12_100
bash dist_train.sh configs/voc12/vpt_seg_zero_vit-b_512x512_20k_12_10.py Path/to/voc12/zero_12_10

Training (Transductive):

bash dist_train.sh ./configs/coco/vpt_seg_zero_vit-b_512x512_40k_12_100_multi_st.py Path/to/coco/zero_12_100_st --load-from=Path/to/coco/zero_12_100/iter_40000.pth
bash dist_train.sh ./configs/voc12/vpt_seg_zero_vit-b_512x512_10k_12_10_st.py Path/to/voc12/zero_12_10_st --load-from=Path/to/voc12/zero_12_10/iter_10000.pth

Training (Fully supervised):

bash dist_train.sh configs/coco/vpt_seg_fully_vit-b_512x512_80k_12_100_multi.py Path/to/coco/fully_12_100
bash dist_train.sh configs/voc12/vpt_seg_fully_vit-b_512x512_20k_12_10.py Path/to/voc12/fully_12_10

Inference:

python test.py ./path/to/config ./path/to/model.pth --eval=mIoU

For example:

CUDA_VISIBLE_DEVICES="0" python test.py configs/coco/vpt_seg_zero_vit-b_512x512_80k_12_100_multi.py Path/to/coco/zero_12_100/latest.pth --eval=mIoU

Cross Dataset Inference:

CUDA_VISIBLE_DEVICES="0" python test.py ./configs/cross_dataset/coco-to-voc.py Path/to/coco/vpt_seg_zero_80k_12_100_multi/iter_80000.pth --eval=mIoU
CUDA_VISIBLE_DEVICES="0" python test.py ./configs/cross_dataset/coco-to-context.py Path/to/coco/vpt_seg_zero_80k_12_100_multi/iter_80000.pth --eval=mIoU

Related Assets & Acknowledgement

Our work is closely related to the following assets that inspire our implementation. We gratefully thank the authors.

• CLIP: https://github.com/openai/CLIP
• Maskformer: https://bowenc0221.github.io/maskformer
• Zegformer: https://github.com/dingjiansw101/ZegFormer
• zsseg: https://github.com/MendelXu/zsseg.baseline
• MaskCLIP: https://github.com/chongzhou96/MaskCLIP
• SegViT: https://github.com/zbwxp/SegVit
• DenseCLIP: https://github.com/raoyongming/DenseCLIP/blob/master/segmentation/denseclip
• Visual Prompt Tuning: https://github.com/KMnP/vpt

Citation:

If you find this project useful, please consider citing:

@article{zhou2022zegclip,
  title={ZegCLIP: Towards adapting CLIP for zero-shot semantic segmentation},
  author={Zhou, Ziqin and Lei, Yinjie and Zhang, Bowen and Liu, Lingqiao and Liu, Yifan},
  journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2023}
}