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.
-
Clone this repository.
git clone https://github.com/turingmotors/openlenda.git cd openlenda -
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 -
Download the pretrained weights.
wget https://github.com/turingmotors/openlenda/releases/download/v0.1.0/openlenda_s.pth -P models
If you don't have a GPU, please remove --device 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 --device gpupython 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.
| Model | size | AP50 | AR50 | Download |
|---|---|---|---|---|
| OpenLenda-nano | 416x416 | 0.941 | 0.830 | link |
| OpenLenda-tiny | 416x416 | 0.944 | 0.807 | link |
| OpenLenda-s | 640x640 | 0.964 | 0.946 | link |
| OpenLenda-x | 640x640 | 0.950 | 0.961 | link |
Please see YOLOX's README or docs (almost same as YOLOX's).
Check Q and A.
Sorry, but the dataset is private.
About 44,000 images.
This is possible under the Apache 2.0 license. However, at this time, it is not possible to use the code in this repository as-is for training and evaluation, so you will need to modify the code yourself.
This project is licensed under the Apache License 2.0. See LICENSE for more details.
This project is based on YOLOX.