readme.md

Uncertainty-aware Object Detection for traffic light

---

PyTorch + MMDetection backbone

Train

Firstly, you need to convert original GTSDB dataset into MS-COCO dataset format.

Put your dataset GTSDB into ../data/ and rename it as original_GTSDB.

Then run

python gtsdb2coco.py

Then run on single GPU

cd codes
CUDA_VISIBLE_DEVICES=X python main.py configs/xxx.json 

Or you can choose launch on multiple GPUs. You can use the template from current sh dist_train_faster_rcnn_r50_2x.sh

CUDA_VISIBLE_DEVICES=6,7 python -m torch.distributed.launch --nproc_per_node=2 --master_port=20001 ./train.py configs/faster_rcnn/faster_rcnn_r50_fpn_2x_gtsdb.py --resume-from /data2/chenpj/UOD/codes/work_dirs/faster_rcnn_r50_fpn_2x_gtsdb/epoch_36.pth  --launcher pytorch

Inference

For inference, just use demo.py !

python demo.py --input_image ${IMAGE} --input_video ${VIDEO} --output {OUTPUT_FILE} --load_from ${MODEL_PATH}

Now, enjoy your traffic light detecting playground!

Model Zoo

We only test Faster-RCNN-FPN-ResNet50-FPN as baseline. Here, we provide the .json config and pre-trained model.

You can use any model in MMDetection to implement the traffic light detecting easily!

backbone	config	pre-trained model	mAP
Faster-RCNN-ResNet50-FPN	config	pre-trained weights	0.5210
Faster-RCNN-ResNeSt50-FPN-SyncBN	TODO	TODO	0.5660

Next Plan

We should add custom uncertainty-head in model and uncertainty-aware loss function.

Acknowledgement

Thanks for open-mmlab/mmdetection: OpenMMLab Detection Toolbox and Benchmark (github.com) to share such a wonderful tool for universal object detection pipeline. If you are beneficial from this repo, please star it :)