This project implements the DeepLab V3 architecture for the semantic image segmentation task trained and evaluated using the Cityscapes dataset. It uses atrous convolutions in both the ResNet-101 backbone architecture and atrous spatial pyramid pooling layer. This implementation has been built using the PyTorch framework.

In order to run the project, use the following command:

python main.py <dataset-folder> <model-checkpoint-folder> <losses-folder> --is-training-model

In order to evaluate the project, use the following command:

python main.py <dataset-folder> <model-checkpoint-folder> <losses-folder> --is-evaluating-model --checkpoint-path <checkpoint-path>

Note that in both cases, the three required arguments are defined as follows:

1. dataset-folder: The path of the Cityscapes dataset
2. model-checkpoint-folder: The path to save model checkpoints every 5 epochs. This can be changed using the --save-checkpoint-index option
3. losses-folder: The path to save the average training and validation losses every 5 epochs. This can be changed using the --save-checkpoint-index option

In this project, I used a momentum of 0.9, weight decay of 5e-4, and a polynomial learning rate scheduler as defined in the paper. The model has been trained for 65 epochs using the learning rate of 1e-2. Then, the last 5 epochs has been trained using a learning rate of 1e-3.

In this project, I am able to achieve a mean intersection over union score of 0.700.

Since the model checkpoint file is quite large of a size, I have placed it in a public Dropbox folder found here. In order to load this, place this inside <model-checkpoint-folder> and pass the filename without the extension to the --checkpoint-path option.

Here is an example of a mask produced by the model and the original image:

Mask image:

Original image: