Predict mask detection with detectron2

The FSOFT organizing committee provides a dataset of 790 pictures labeled with bounding boxes in 3 classes: wearing masks, not wearing masks and wearing masks incorrectly.

Create virtual environment

conda create -n d2detector python=3.8
conda activate d2detector

Install requirements

Torch CUDA and Detectron2

pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu111/torch1.10/index.html

Include Detectron2: Follow CUDA and Torch version to run this project.

Link detail: https://detectron2.readthedocs.io/en/latest/tutorials/install.html#install-pre-built-detectron2-linux-only

Related packages

pip install -r requirements.txt

Dataset

• mask-fpt-ai.zip: Include 1 folder images (790 images) and 1 annotations coco format
• annotations.json: Annotations of above dataset
• train.json: Annotations for training (100%)
• test.json: Annotations for testing (20%)

Get data (Sample)

bash setup_data.sh

Include: 1 folder images (790 images) and 1 annotations coco format

Deal video dataset

I have already define some functions to convert dataset video to images.

python d2detector/video2frame.py --data_paths /path/to/dataset/video --output dataset/images

Split train test

python d2detector/cocosplit.py --having-annotations -s 0.9 dataset/annotations.json dataset/train.json dataset/test.json

Train

python train.py \
    --datapath_train 'dataset/train.json' \
    --datapath_test 'dataset/test.json' \
    --imagepath 'dataset/images'

# or simply
python train.py

Inference

python inference.py \
    --input_path 'sample/978.jpg' \
    --model_zoo 'COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml' \
    --weight 'output/model_final.pth' \
    --threshold 0.7 \
    --fps 0 \
    --output_path 'inference/' 

# or simply
python inference.py

With fps:

• 0 is image
• 1 is video
• greater than 1 is predict frame per second on video

Check image or video in inference folder

Output

Sample

Export model to tracing

python d2detector/export_model_tracing.py \
    --sample_image 'sample/978.jpg' \
    --model_path 'output/model_final.pth' \
    --threshold 0.7 \
    --output_path 'output/' 

After have model tracing, you can follow my repository about inference Detectron2 by C++.

Link repo: https://github.com/vnk8071/AI-on-Cpp/tree/main/Detectron2-Cpp

License

@misc{wu2019detectron2,
  author =       {Yuxin Wu and Alexander Kirillov and Francisco Massa and
                  Wan-Yen Lo and Ross Girshick},
  title =        {Detectron2},
  howpublished = {\url{https://github.com/facebookresearch/detectron2}},
  year =         {2019}
}