IEEE BigData 2022

Crowdsensing-based Road Damage Detection Challenge (CRDDC2022)

Team name: T22_033_Dai_Quoc_Tran

1. Prepare data

In prepare_data, using these notebooks to convert data format into mmdetection and yolov5 format.

2. Train model

MMDet

Clone mmdetection, follow tutorial for installing. Using vfnet_config to train.

Yolov5

Clone Yolov5, follow tutorial for installing. Using road_crack_coco.yaml to train.

3. Inference

Trained weights

Download train weights. PWD: crddc2022.

MMDet

Norway, with its high-resolution rectangular image size, works well with VFNet. Using inference_norway to visualize and prepare for submission.

Yolov5

Other countries with low-resolution images and square image sizes work well with Yolov5. Using inference_all to inference other countries and ensemble for overall countries submission.

4. Ensemble and submission

We can adjust the hyper-parameter for ensemble_boxes in the inference files to improve accuracy.