CamVid Segmentation

Deep Learning-based Semantic Segmentation for Autonomous Driving

This project was developed as a part of the presentation that I gave on the Programming 2.0 webinar: Autonomous driving.

The presentation slides can be found here: programming_2_0.pdf

Instructions

To be able to use the code please follow listed instructions:

1. Download data from https://www.kaggle.com/carlolepelaars/camvid

2. Extract files and place them into data folder using the following folder structure:

   data/test/*.png
   data/test_labels/*.png
   data/train/*.png
   data/train_labels/*.png
   data/val/*.png
   data/val_labels/*.png
   data/class_dict.csv
   

3. Execute train.py to train the model based. The results will be placed in tmp folder. In case of an out of memory problem, adjust batch size in settings.py:

   batch_size = 4
   

4. Execute evaluate.py to evaluate trained model on the test subset. The results will be placed in tmp folder.

5. Check settings.py for other training options.

Project Overview

Please overview UML diagram that depicts major project components and dependencies:

The project have two executable scripts:

• train.py - used to train a model.
• evaluate.py - used to evaluate trained model against test set.

There is also several utility scripts with the following responsibilities:

• settings.py - add project related settings.
• data.py - utility variables, functions, and classes responsible for accessing dataset images.
  • Relies on tensorflow and cv2 libs.
• model.py - utility functions used to create segmentation model.
  • Relies on myresunet.py script and segmentation_models lib.
• myresunet.py - utility functions used to create custom U-Net/ResNet inspired model.

For more info about segmentation_models lib please check its GitHub page: https://github.com/qubvel/segmentation_models

Experiments

A list of conducted experiments with appropriate accuracy achieved on the test set is reported in experiments.csv.

The best accuracy of 90.05% (min. 70.86%, max. 97.40%, std. 5.68%) were achieved with U-Net model and EfficientNetB2 backbone. The trained model along with all other outputs for that experiment can be found here.