Count Anything
Liqi, Yan
ZJU-CV, Zhejiang University / Fudan Univerisity

Count Anything (CA) project is a versatile image processing tool that combines the capabilities of Segment Anything, Semantic-Segment-Anything, and CLIP. Our solution can count any object specified by users within an image.

The CA engine consists of three steps:

• (I) Segement Anything. Following Semantic-Segment-Anything, CA engine crops a patch for each mask predicted by Segment Anything.
• (II) Class Mixer. To identify the masks that match the user’s text prompt, we add the text prompt as an additional class into the class list from the close-set datasets (COCO or ADE20K).
• (III) CLIP Encoders. The CA engine uses CLIP image encoder and text encoder to assess if the text prompt is the best option among other classes. If yes, this mask is considered as an instance of the class given by the text prompt, and the count number is incremented by 1.

• Python 3.7+
• CUDA 11.1+

conda env create -f environment.yaml
conda activate ca-env

Please use --convert-to-rle to save segmentation results as .json files.

python scripts/amg.py --checkpoint sam_vit_h_4b8939.pth --model-type vit_h --convert-to-rle --input examples/AdobeStock_323574125.jpg --output output --pred-iou-thresh 0.98 --crop-n-layers 0 --crop-nms-thresh 0.3 --box-nms-thresh 0.5 --stability-score-thresh 0.7

python scripts/amg.py --checkpoint sam_vit_h_4b8939.pth --model-type vit_h --convert-to-rle --input examples/crowd_img.jpg --output output --pred-iou-thresh 0 --min-mask-region-area 0  --stability-score-thresh 0.8

├── Count-Anything
|   ├── data
|   │   ├── examples
|   │   │   ├── AdobeStock_323574125.jpg
|   │   │   ├── AdobeStock_323574125.json
|   │   │   ├── ...

Please use --text_prompt [OBJ] to specify the customized class for counting.

python scripts/main.py --out_dir=output --world_size=1 --save_img --text_prompt="shirt" --data_dir=data/examples 

python scripts/main.py --out_dir=output --world_size=1 --save_img --text_prompt="person" --data_dir=data/crowd_examples/ 

The result is saved in output folder.

• Segment Anything provides the SA-1B dataset.
• HuggingFace provides code and pre-trained models.
• Semantic-Segment-Anything provides code.
• CLIP provide powerful semantic segmentation, image caption and classification models.

If you find this work useful for your research, please cite our github repo:

@misc{yan2023count,
    title = {Count Anything},
    author = {Yan, Liqi},
    howpublished = {\url{https://github.com/ylqi/Count-Anything}},
    year = {2023}
}