container_ocr

• This project use DBNet to detect text and SVTR to recognize text and K-means to cluster text.

• The simplest way to run this project is to use the docker image.

• You can build the image following the instructions in the Dockerfile.

• Currently, the project infer on GPU and requires CUDA >= 11.7.

• I only test recognition on video.

• The inference speed is about 25FPS on a single RTX 2080Ti with frame size: 1280x720.

• Especially, You must put pretrained model and put in the folder ./assets/weights and video sample in the folder ./assets/images before running the project.

• To run this project, you can use the following command:

  docker build -t container_ocr .

  export DISPLAY="IP_HOST:0.0", replace IP_HOST with your IP address.

  xhost +

  docker run -it --rm --runtime=nvidia --gpus all -v /path/to/your/project/assets:/workspace/assets -e DISPLAY=$DISPLAY container_ocr bash

  mkdir build && cd build

  cmake ..

  make -j8

  ./container_ocr ../assets/images/sample.mp4

• This project is not developed for production, so it is not optimized for speed.

• This project can only run on Libtorch and Torchvision in C++ 17 because of limitation of some operartors. Thus, I cannot export to ONNX format and run on TensorRT or OpenVino.

• You can download pretrained model at here