This repo contains an implementation of the FCN-8 neural network architecture to label road pixels in images.
The model in main.py was trained over 30 epochs with a batch size of 10 images, which reduced the average loss to approximately 0.034.
Some sample output on the test images is shown below. Also the model can be applied to video frames as demonstrated here.
The model often mis-classifies pixels when the lighting conditions vary significantly from the training data. Possible approaches to mitigate this could include:
- Pre-processing images to improve their quality before applying the labelling algorithm; and/or
- Augmenting the training data by including images with artificially reduced/increased brightness.
