Overview / Usage

The state of the roads plays a huge role in avoiding accidents which endangers people's lives, which is why maintenance and management of roads should be done periodically and exhaustively. However, due to lack of financial and qualified human resources, many local governments are not able to conduct sufficient inspections.
In Japan for example, some municipalities automate damage determination by using high-performance sensors, but because of their high cost, many municipalities are unable to introduce them. Therefore, there is a need for a method that makes it easy to judge the damage of the road surface at low cost.

This is what this challenges is trying to tackle. By using vehicle-mounted smartphone we get a photograph of road images and then we use this collected images to train our model to detect the damages and classify them.

When this challenge goes well, there is a possibility that a simple road check can be done using only smartphones.

The training and test data consist of 9,053 photographs, collected from smartphone cameras, hand labeled with the presence or absence of 8 road damage categories.

For this challenge, a random subset of 7,240 of the images with labels where provided as the training set along with a list of the 8 categories. The remaining images will be used as the test set. The test data for this competition are not contained in the training data provided.

Methodology / Approach

1- State of the art model used :

This challenge consists of detecting and classifying road damage images that is why , after a throughout research, I decided to use a deep learning ,computer vision state of the art model which is SSD.

SSD stands for Single Shot Detection, a MultiBox object detection technique, its paper : Single Shot MultiBox Detector (by C. Szegedy et al.) was released at the end of November 2016 and reached new records in terms of performance and precision for object detection tasks, scoring over 74% mAP (mean Average Precision) at 59 frames per second on standard datasets such as PascalVOC and COCO.

Here is a simple explanation about what each term means, taken from ' Understanding SSD MultiBox — Real-Time Object Detection In Deep Learning, Towards Data Science written by Eddie Forson' :

• Single Shot: this means that the tasks of object localization and classification are done in a single forward pass of the network
• MultiBox: this is the name of a technique for bounding box regression developed by Szegedy et al.
• Detector: The network is an object detector that also classifies those detected objects

SSD-500 (the highest resolution variant using 512x512 input images) achieves best mAP on Pascal VOC2007 at 76.8%, but at the expense of speed, where its frame rate drops to 22 fps. SSD-300 is thus a much better trade-off with 74.3 mAP at 59 fps.
Which is why I choose to train my model on SSD-300.

2- The base network for the model:

This was for the state of the art model, as for the base network I am training simultaneously using VGG ILSVRC dataset of 16 layers reduced and VGG16 trained on the COCO dataset. Taken from this repository: https://github.com/pierluigiferrari/ssd_keras.

3- The open source library:

I am currently using Keras which is an open source neural network library written in Python to develop my model.

Technologies Used

• Keras with Tensorflow backend: as a python open source neural network library.
• Intel AI DevCloud: as the cloud where I train my model.

Repository

https://drive.google.com/drive/folders/16uTTaUb4yoASx1X4iYMJwXP_Rb56v4cl?usp=sharing

Other links

• Road Damage Detection Challenge website

• Github repo containing the pieces of code I worked with (https://github.com/pierluigiferrari/ssd_keras)

• Understanding SSD MultiBox 
• Road Damage Detection Using Deep Neural Networks with Images Captured Through a Smartphone (Original Paper)