Ready to deploy blood cell detection model for NVIDIA Jetson platform on DeepStream SDK. Trained with using TAO Toolkit based on DetectNet V2 pre-trained model. The model can detect RBCs and WBCs from peripheral blood smear.
- NVIDIA Jetson Developer Kits or Production Module Systems
- CSI Camera, USB Camera or h264 Encoded Video
- Ubuntu 18.04 L4T
- NVIDIA JetPack SDK 4.6.1
- NVIDIA DeepStream SDK 6.0.1
I used blood cell detection dataset which I collected and annotated when I was in medical school. Visit the dataset repository for further information.
- Be sure that you properly installed JetPack SDK 4.6.1.
- Install NVIDIA DeepStream SDK 6.0.1. You can follow the official documentation.
- Clone DeepStream TAO Apps repository. Notice that we cloned the branch for DeepStream 6.0.1 version. an
git clone -b release/tao3.0_ds6.0.1 https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps/
- Open deepstream_tao_apps directory, set CUDA version variable and compile.
cd deepstream_tao_apps
export CUDA_VER=10.2
make
- Clone blood-cell-detection-model repository.
git clone https://github.com/draaslan/blood-cell-detection-model
- Run ds-tao-detection app with config and test video. This process can be take a while.
./apps/tao_detection/ds-tao-detection -d \
-c blood-cell-detection-model/inference_config.txt \
-i blood-cell-detection-model/test.h264
- Model output will be saved as h624 format and inference result will be open on a window.
- Add video and image demonstration.
- Train model using more data with augmentation.
- Improve inference config.
- Containerize full application.
See LICENSE for details.