Road-Crack-Detection-Using-Instance-Segmentation (Roads360 v.2)

This is the Repository for

Final Year Capstone Thesis Project: A Complete Road Health Monitoring System for Bangladesh: Crowdsourced Data Collection through User Application and Smart Analyzer with Real-time Feedback System for Driving Assistance.

Codes and Model Credit Goes to

• YOLACT: Real-time Instance Segmentation
• YOLACT++: Better Real-time Instance Segmentation

My Previous work on Semantic Segmentaion Roads360

---

Objective

Bangladesh has been fighting the issue of ensuring a safer road management system since its inception. Yet Bangladesh is suffering from an incapable and unattended road communication system. Building good-quality roads with sustainable materials covers only a part of maintaining and preserving a safe road system. But maintaining real-time observation and constant data collection about the routes is a must to ensure safe roads for safe commutation. Automation in real-time data collection about the road surface condition and providing analyzed feedback to the government and people can effectively reduce road accidents across the country. In this project, we present a smart system for road health monitoring. The two main branches of this project are Large-scale crowd-sourced data collection of road-surface condition through user application and Real-time feedback based on image segmentation to detect road-cracks and anomalies. We believe that such a crowd-sourcing-based platform will be highly effective in ensuring a safer commutation experience for citizens of Bangladesh. In autonomous driving in Bangladesh, given a front camera view, the car needs to know where the road is and Crackers on the road surface. In this project, we trained a neural network to label the pixels of a road in images by using a Model YOLACT and trained with our Own collected images for road and Crack segmentation.

Watch the Demo video : here

- "Road" = Colored in Red

499 Capstone Project Showcase Poster

Installation

• Clone this repository:

  git clone https://github.com/sajidahmed12/Road-Crack-Detection-Using-Instance-Segmentation-Roads360-v.2
  cd Road-Crack-Detection-Using-Instance-Segmentation-Roads360-v.2

• Set up the environment using one of the following methods:
  • Using Anaconda
  • Manually with pip
    • Set up a Python3 environment (e.g., using virtenv).
    • Install Pytorch 1.0.1 (or higher) and TorchVision.
    • Install some other packages:

      # Cython needs to be installed before pycocotools
      pip install cython
      pip install opencv-python pillow pycocotools matplotlib 

1.2 Dependencies & my environment

Anaconda is used for managing the environment.

• Python 3.7.*+ , Pytorch v1.1, CUDA 10.0 CuDnn 7.5.0, Numpy, SciPy,glob,tdqm
• OS: Windows 10 Pro
• CPU: Intel® Core™ i7-7700K CPU @ 4.00-4.60 GHz 4 core × 8 Threads
• GPU: NVidia GeForce GTX 1070 (8GB GDDR5X VRAM)
• RAM: 16GB DDR4 2400 MHz

Yolact Weights

YOLACT++ models default weights:

Image Size	Backbone			Weights
550	Resnet50-FPN			yolact_plus_resnet50_54_800000.pth	Mirror
550	Resnet101-FPN			yolact_plus_base_54_800000.pth	Mirror

To evalute the model with the Yolact base default config , put the corresponding weights file in the ./weights directory and run one of the following commands. The name of each config is everything before the numbers in the file name (e.g., yolact_base for yolact_base_54_800000.pth).

Our Trained Model Weights

Model weights:

Image Size	Backbone			Weights
550	Yolact++ with Resnet101-FPN			yolact_base_23688_1066000.pth	Mirror

|

To evalute the model with our trained Model for Road/Crack Segmentation, put the corresponding weights file in the ./weights directory and run one of the following commands. The name of each config is everything before the numbers in the file name (e.g., yolact_base for yolact_base_54_800000.pth).

Testing the Model (with image)

# Results on the specified image.
python eval.py --trained_model=weights/yolact_base_23688_1066000.pth --image=my_image.png

# Process an image and save it to another file.
python eval.py --trained_model=weights/yolact_base_23688_1066000.pth --image=input_image.png:output_image.png

# Process a whole folder of images.
python eval.py --trained_model=weights/yolact_base_23688_1066000.pth --images=path/to/input/folder:path/to/output/folder

Testing the Model (with video)

# Display a video in real-time. Author's Suggests that "--video_multiframe" will process that many frames at once for improved performance.
# If you want, use "--display_fps" to draw the FPS directly on the frame.
python eval.py --trained_model=weights/yolact_base_23688_1066000.pth --video_multiframe=4 --video=my_video.mp4

# Display a webcam feed in real-time.
python eval.py --trained_model=weights/yolact_base_23688_1066000.pth  --video_multiframe=4 --video=0

# Process a video and save it to another file.
python eval.py --trained_model=weights/yolact_base_23688_1066000.pth --video_multiframe=4 --video=input_video.mp4:output_video.mp4

for more try the eval.help

python eval.py --help

Training

We trained on with out own dataset. Make sure to download the entire dataset from the link here.

• To train, grab an pretrained model and put it in ./weights folder.

  • Download yolact_base_23688_1066000.pth from here

• Run one of the training commands below.

  • Note that you can press ctrl+c while training and it will save an *_interrupt.pth file at the current iteration.
  • All weights are saved in the ./weights directory by default with the file name <config>_<epoch>_<iter>.pth.

# Trains using the base config with a batch size of 8 .
python train.py --config=yolact_base_config

# Trains yolact_base_config with a batch_size of 5. For the 550px models, 1 batch takes up around 1.5 gigs of VRAM, so specify accordingly.
python train.py --config=yolact_base_config --batch_size=5

# Resume training yolact_base with a specific weight file and start from the iteration specified in the weight file's name.
python train.py --config=yolact_base_config --resume=weights/yolact_base_10_32100.pth --start_iter=-1

# Use the help option to see a description of all available command line arguments
python train.py --help

Custom Datasets

We have used this method for training our own dataset by following these steps:

• We have created a definition for our dataset under dataset_base in data/config.py

ROAD_CLASSES = ("road", "crack", "surroundings")

road_dataset = dataset_base.copy({
    'name': 'Road Dataset',

    'train_images': 'data/data_road/training',
    'train_info':   'data/data_road/training/annotations.json',

    'valid_images': 'data/data_road/training',
    'valid_info':   'data/data_road/training/annotations.json',

    'has_gt': True,
    'class_names': ROAD_CLASSES
})

SOME RESULTS

References

• 1 Bolya, D., Zhou, C., Xiao, F., & Lee, Y. J. (2019). Yolact: Real-time instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 9157-9166).

if any issues questions please please contact Md Sajid Ahmed