Awesome

Detect-AnyShadow

Detect-AnyShadow (DAS) is a tool for detecting shadows in videos. With very little user interaction, DAS can detect shadows throughout the whole video. It is built upon segment anything model (SAM) 「Paper」 and we have fine-tuned SAM to enable it to detect shadows. Then we leverage a long short-term network (LSTN) to enable the network for video shadow detection.

Demo Tutorial

https://github.com/harrytea/Detect-AnyShadow/assets/44457717/c31fd723-3867-4106-a30c-fd477a10027d

Run python demo_app.py and refer to the operation of the upper demo.

# use 1 lsab block
python demo_app.py --model lstnt --ckpt_path ./lstn/checkpoints/lstnt --start_step 10000

# use 2 lsab block
python demo_app.py --model lstns --ckpt_path ./lstn/checkpoints/lstns --start_step 10000

# use 3 lsab block
python demo_app.py --model lstnb --ckpt_path ./lstn/checkpoints/lstnb --start_step 10000

Note: Gradio does not support drawing bounding boxes during the user interaction (for more information, please refer to issue). Therefore, we convert the connected regions that the user has marked into minimal bounding boxes in the code.

:tada: News

2023/07/12: Update the fine-tuning code of SAM. Update the training and testing code of LSTN.
2023/05/27: We released Detect-AnyShadow publicly, updated a simple version of the online demo, and provided test code for the ViSha dataset.

:wrench: Pretrained Models

Download the original SAM ViT-B here (code: jx90), and put it into folder './checkpoint'

Download the fine-tuned SAM ViT-B here (code: wons), and put it into folder './checkpoint/chk_sam'

Download the lstnt (code: 0ml4), lstns (code: obub), lstnb (code: 2e18) and put it into folder './lstn/checkpoints'

:sunny: Get Started

1. Installation

conda create -n das python=3.10
pip install -r requirements.txt

2. Finetune SAM

python sam_finetune.py

Before running, modify the training_path to your own visha path. Download the sam_vit_b_01ec64.pth and put it to ./checkpoints folder. The checkpoints will be saved in ./checkpoints/chk_sam.

3. Test SAM

python sam_test.py

Before running, modify the path to your own visha path. The results will be saved in ./results/sam folder

4. Train LSTN

CUDA_VISIBLE_DEVICES='0,1,2,3' python lstn/tools/train.py --amp --exp_name lstn --stage visha --model lstnb --gpu_num 4 --batch_size 16

Before running, go to the lstn/configs/default.py file and modify self.DIR_VISHA param to your own visha data path. The checkpoints will be saved in ./checkpoints/lstn_xxxx folder. The terminal output is as follows:

I:0, LR:0.00002, Ref(Prev): L 0.562 IoU 15.0%, Curr1: L 0.560 IoU 12.4%, Curr2: L 0.571 IoU 11.1%, Curr3: L 0.569 IoU 11.1%, Curr4: L 0.556 IoU 13.8%
I:20, LR:0.00002, Ref(Prev): L 0.392 IoU 43.7%, Curr1: L 0.376 IoU 45.3%, Curr2: L 0.371 IoU 45.4%, Curr3: L 0.370 IoU 47.4%, Curr4: L 0.373 IoU 46.4%
I:40, LR:0.00002, Ref(Prev): L 0.248 IoU 58.9%, Curr1: L 0.263 IoU 60.5%, Curr2: L 0.254 IoU 62.0%, Curr3: L 0.253 IoU 64.2%, Curr4: L 0.251 IoU 63.6%
I:60, LR:0.00002, Ref(Prev): L 0.232 IoU 65.8%, Curr1: L 0.231 IoU 68.3%, Curr2: L 0.231 IoU 67.7%, Curr3: L 0.237 IoU 66.5%, Curr4: L 0.244 IoU 63.0%
I:80, LR:0.00002, Ref(Prev): L 0.205 IoU 69.2%, Curr1: L 0.200 IoU 65.4%, Curr2: L 0.221 IoU 61.7%, Curr3: L 0.217 IoU 63.3%, Curr4: L 0.216 IoU 63.8%
I:100, LR:0.00002, Ref(Prev): L 0.145 IoU 76.6%, Curr1: L 0.164 IoU 76.3%, Curr2: L 0.161 IoU 74.9%, Curr3: L 0.168 IoU 77.6%, Curr4: L 0.139 IoU 77.4%
...

Here, we preloaded the dataset using numpy. If you are using original jpg or png format images, you will need to make slight modifications to the lstn/dataloaders/train_datasets.py.

5. Test LSTN

python lstn/tools/test.py --stage visha --exp_name lstn --model lstnb --datapath your_data_path --start_step 1000 --ckpt_path your_chk_path

For test ViSha Dataset, put the testing data in dataset, list the data as follows, and run python test.py. The shadow mask will be saved in results/lstn_xxxx/xxx.

- test
    - images
        - video1
            - xx1.jpg
            - xx2.jpg
            - xx3.jpg
            ...
    - labels
        - video1
            - xx1.png
        - video2
            - xx2.png

our prediction results (code: uhaq) is here.

6. Inference

python ./lstn/eval/evaluate.py --epoch=10000 --pred_path="./results/xxx"

:book: Citation

If you find our work useful in your research or applications, please cite our article using the following BibTex.

@article{wang2023detect,
  title={Detect any shadow: Segment anything for video shadow detection},
  author={Wang, Yonghui and Zhou, Wengang and Mao, Yunyao and Li, Houqiang},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2023},
  publisher={IEEE}
}

:blush: Acknowledge

The code for this project brings heavily from the following repositories. We thank the authors for their great work.

facebookresearch: segment-anything gaomingqi: Track-Anything z-x-yang: Segment-and-Track-Anything yoxu515: aot-benchmark