Road Damage Detection Challenge (IEEE Big Data Cup 2020)

• D2_rdd2020_test.py : Runs the test on dataset test1/* or test2/* when provided with config and model weights in the codebase. It also writes the submittable output file.
• D2_rdd2020.py : Training codebase for Detectron2 models
• data_rdd.py : Data utility for loading and transforming the Road damage data to COCO format and create train-val-test split
• requirements : use this file to install packages in a python virtual environment. Use command "pip install -r requirements"

• 1.) Create a basic toolbox using
  • Detectron2
• 2.) Refer to old dataset/model for augmentation/init at https://github.com/sekilab/RoadDamageDetector
• Resnet 50, 101 and various Hyper parameters in Faster-R-CNN model
• RetinaNet and YoloV5
• Mix train & test dataset and fine tune
• Mix train & val dataset and fine tune

• 1.) Setup Data split for train/val/test using [data_rdd.py]

  • Add the train/ data in a folder called rdd2020/
  • Set the variable in data_rdd.py(line 10) based on the path of your dataset > DETECTRON2_DATASETS = "/media/rahul/Karmic/data"
  • Expect a soft link of train files in folder rdd2020/lval, rdd2020/ltrain and rdd2020/ltest with annotations
  • Manually Merge rdd2020/lval and rdd2020/ltrain to create rdd2020/lvaltrain dataset for training
  • Current dataset in code is configured for training on 'rdd2020/lvaltrain' data

• 2.) Configure the training codebase using [D2_rdd2020.py]. Two strategies are shown below,

  • Config For best Test2 score training
    • Update relevant line:130-154 with the following configurations

        cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml"))
        cfg.MODEL.WEIGHTS         = model_zoo.get_checkpoint_url("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml")
        cfg.SOLVER.BASE_LR        = 0.01                
        cfg.SOLVER.MAX_ITER       = 30000               
        cfg.SOLVER.STEPS          = (23000, 25000, 26000)
        cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE  = 640
      

  • Config For best Test1 score training
    • Update relevant line:147-154 with the following configurations

        cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_101_FPN_3x.yaml"))
        cfg.MODEL.WEIGHTS         = model_zoo.get_checkpoint_url("COCO-Detection/faster_rcnn_R_101_FPN_3x.yaml")
        cfg.SOLVER.BASE_LR        = 0.015               
        cfg.SOLVER.MAX_ITER       = 30000               
        cfg.SOLVER.STEPS          = (25000, 28000)
        cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE  = 4096
      

• 3. Train by running command 'python D2_rdd2020.py'

  • upon completing the max iterations a 'model_final.pth' will be dumped in respective output folder.

• 4. Test the dumped model_final.pth weights by using [D2_rdd2020_test.py]

  • Pre-trained model weights download at https://github.com/vishwakarmarhl/rdd2020/releases/tag/b0.1
  • Configure the line:32 for the model configuration based on the training strategy used in Step 2
  • Comment or uncomment line:49 (test1) or line:50 (test2) based on the test dataset you want to generate the submission file for.
  • run the test command 'python D2_rdd2020_test.py' to find a txt dump file that evaluates the test images using the trained model
  • use this txt dump file for submission or eval purposes