OpenLenda

OpenLenda is a Japanese traffic light detection and recognition model.

OpenLenda is a Japanese traffic light recognition model based on YOLOX. This model not only detects traffic lights but can also recognize their colors and specific arrow indications in an end-to-end manner. For more information about the development, please visit https://zenn.dev/turing_motors/articles/traffic-light (Japanese article). This repository is for model publication.

1. Clone this repository.

   git clone https://github.com/turingmotors/openlenda.git
   cd openlenda

2. Install python dependencies. We strongly recommend using pyenv and poetry to manage python versions and dependencies.

   pyenv install 3.8.10
   pyenv local 3.8.10
   poetry install

3. Download the pretrained weights.

   wget https://github.com/uringmotors/openlenda/releases/latest/download/openlenda_s.pth -P models

If you don't have a GPU, please remove --gpu from the following commands.

python tools/demo.py video -f exps/openlenda_s.py\
    -c models/openlenda_s.pth\
    --path assets/tokyo_day.mp4\
    --conf 0.5 --nms 0.01 --tsize 640 ---save_result --gpu

python tools/export_onnx.py --output-name openlenda_s.onnx\
    -f exps/openlenda_s.py\
    -c models/openlenda_s.pth\

AP and AR are evaluated on the private validation set. The results are shown below.

Please see YOLOX's README or docs (almost same as YOLOX's).

Turing Inc.

This project is licensed under the Apache License 2.0. See LICENSE for more details.

This project is based on YOLOX.