Waynet - A Road Segmentation project

Author

• Vishal Adithya.A

Overview

This model demonstrates a road segmentation implemented using deep learning techniques which predicts the road regions in the input image and returns it in a grayscale format.

Models

• rs1-low.pth: The lowest performer model with a loss of 0.69%.
• rs1-high.pth: The highest performer model with a loss of 0.07%.

Model Structure

Features

1. Architecture

   • Basic Resnet50 model with few upsampling and batch normalisation layers.
   • Contains over 60 million Trainable paramameters.
   • Training Duration: 1 hour.

2. Training Data

   • Source: (bnsapa/road-detection)
   • Format: The dataset includes RGB images of roads around the globe and their corresponding segment and lane.
   • Preprocessing: With the help of torch and torchvission api basic preprocessing like resizing and convertion to tensor were implemented.

3. CostFunctions Score

   • BCE: 0.07
   • MSE: nil
   • [NOTE: All the above scores are trained using the highest performer model]

License

This project is licensed under the Apache License 2.0.

Acknowledgments

• Apple M1 Pro 16gb of unified memory for efficient GPU acceleration during model training
• Pytorch for robust deep learning framework