northautumn / tf-faster-rcnn-rddc

Road Damage Detection and Classification with Faster R-CNN (BigData 2018)

Road Damage Detection and Classification with Faster R-CNN

Introduction

This repository contains source files of Road Damage Detection and Classification (RDDC) based on Faster R-CNN, which achieved a Mean F1-Score of 0.6255 in the Road Damage Detection and Classification Challenge that held as one of the 2018 IEEE Big Data Cup and won the Silver Prize (Ranked 2nd).

It is developed based on the tf-faster-rcnn repository and created by Wenzhe Wang.

Contents

1. Requirements
2. Installation
3. Setup Data
4. Submission
5. Training
6. Analysis
7. Citation

Requirements

1. A GPU with sufficient memory.

2. A basic Tensorflow installation (I use 1.4.0).

Installation

1. Clone the tf-faster-rcnn-rddc repository into $RDDC_ROOT

   git clone https://github.com/zhezheey/tf-faster-rcnn-rddc.git

2. Install python packages you might not have in requirements.txt

   # Python 3.4.3+ is needed
   pip install -r requirements.txt

3. Update your -arch in setup script to match your GPU refer to tf-faster-rcnn

   cd $RDDC_ROOT/lib
   # Change the GPU architecture (-arch) if necessary
   vim setup.py

4. Build the Cython modules

   cd $RDDC_ROOT/lib
   make clean
   make

5. Install the Python COCO API

   cd $RDDC_ROOT/data
   git clone https://github.com/pdollar/coco.git
   cd coco/PythonAPI
   make

Setup Data

1. Download the RDDC dataset and convert it into PASCAL VOC 2007's annotation format. For convenience, you can also download the converted dataset here.

2. Create symlinks for the converted PASCAL VOC dataset

   cd $RDDC_ROOT/data
   ln -s $road_damage_dataset VOCdevkit2007

Submission

Follow the steps below to reproduce the experimental results of our submission.

1. Download our pre-trained model on RDDC and place them inside the $RDDC_ROOT/output/res152/voc_2007_trainval/default directory.

2. Get the submission

   cd $RDDC_ROOT/tools
   # Outputs are saved under $RDDC_ROOT
   python submmision.py

Training

1. Download pre-trained models and weights. The current code supports VGG16 and Resnet V1 models. Pre-trained models are provided by slim, you can get the pre-trained models here and put them under the $RDDC_ROOT/data/imagenet_weights folder. For Resnet152, you can set up like:

   mkdir -p $RDDC_ROOT/data/imagenet_weights
   cd $RDDC_ROOT/data/imagenet_weights
   wget -v http://download.tensorflow.org/models/resnet_v1_152_2016_08_28.tar.gz
   tar -xzvf resnet_v1_152_2016_08_28.tar.gz
   mv resnet_v1_152.ckpt res152.ckpt

2. Augment the training data refer to $RDDC_ROOT/analysis/data_augmentation_balance.ipynb.

3. Train

   cd $RDDC_ROOT
   ./experiments/scripts/train_faster_rcnn.sh [GPU_ID] pascal_voc res152

   Note: Please check that you have deleted the pre-trained models before training.

4. Visualization with Tensorboard

   cd $RDDC_ROOT
   tensorboard --logdir=tensorboard/res152/voc_2007_trainval/ --port=[PORT] &

5. By default, trained networks are saved under:

   $RDDC_ROOT/output/res152/voc_2007_trainval/default/
   

   Tensorboard information is saved under:

   $RDDC_ROOT/tensorboard/res152/voc_2007_trainval/default/
   $RDDC_ROOT/tensorboard/res152/voc_2007_trainval/default_val/
   

Analysis

Some analyses during the competititon are under $RDDC_ROOT/analysis.

Citation

Please consider citing the related papers and linking this project.