Project Road Segmentation

This project focuses on an image binary semantic segmentation problem, which extract roads from a set of satellite images. We implemented a model based on Dinknet152, and also introduced data augmentation schemes and mechanism to fine-tune hyperparameters. We finally reached an accuracy of 0.957 and a F1 score of 0.918, ranking 7th on AIcrowd leaderboard.

Structure of the project

├── Data_of_AICorwd_best_result
│   ├── DinkNet152_AIcrowd_0.957
│   ├── DinkNet152_AIcrowd_0.957.csv
│   └── DinkNet152_AIcrowd_0.957.log
│
├── dataset
│   ├── test_set_images
│   └── train
│       ├── groundtruth
│       └── images
│
├── networks
│   ├── __init__.py
│   └── dinknet.py
│
├── submits
│
├── weights
├── README.md
├── data_preprocessing.py
├── framework.py
├── loss.py
├── mask_to_submission.py
├── run.py
├── test.py
└── train.py

• Data_of_AICorwd_best_result: the folder that contains all the data that we use to get our best result on AIcorwd.DinkNet152_AIcrowd_0.957 is the folder that contains the the prediction images that gain an accuracy of 0.957 on AIcrowd. DinkNet152_AIcrowd_0.957.csv is the file is generated from the DinkNet152_AIcrowd_0.957 using mask_to_submission.py.DinkNet152_AIcrowd_0.957.log is the log file records the training loss and validation loss during the training phase.
• dataset: the folder contains the images for training and testing
• /networks/dinknet.py : contains all the neural network models the program uses. Here we define the LinkNet and DinkNet for training and comparison.
• submits: the folder where the testing prediction will be stored.
• weights: The folder where the optimal weight(.th file) generated during the training phase will be stored. When doing the test, the .th file in this folder will also be used. Given that the size of the .th file that reach our best result on AIcrowd is so large(more than 1G) that it exceeds github's file size limitation, we decide not to upload this file. Please feel free to contact me at [email protected] if you need it.
• README.md: that is this file.
• data_peprocessing.py: the file that performs data preprocessing, including resizing, rotation, scaling,flip, color space transformation, etc.
• framework.py: the file that performs general framework manipulation.
• loss.py: the file that defines the loss function(BCE loss, Dice loss)
• mask_to_submission.py : the file that transfers the resulting pictures in /submits/SpecificName folder to the .csv file.
• run.py: the file that integrate data preprocessing, training, and testing.
• test.py: the file that applies test-time augmentation when generating the result of images in /dataset/test_set_images
• train.py: the file that perfom the train
• logs: a folder that will be newed during the training phase. Log that records the training loss and validation loss will be stored in this folder.

Usage

• You can run this project by using following command:

python run.py

Both of the training and testing are porformed on Google Colab GPU(12G). The training phase takes about 2200 seconds, and the testing takes about an hour.