Semantic Segmentation using FCN and UNet models, done as a part of Assignment 4 of the Computer Vision course (IIIT-Hyderabad, Spring '25). The assignment details can be found in the assignment document.

Results

FCN

For ablations and visualisations, see the report.

Visualisations

Each image has three columns, representing input image, ground truth, followed by prediction (each image is a collage of these three in that order).

Backbone type	32s	16s	8s
Frozen
Unfrozen

To train and test the model, run:

cd FCN
python -m src.main <options>

The options include the data directory, batch size, model variant, frozen vs unfrozen backbone, the number of epochs, and the learning rate. See the main.py file for more details.

To visualise some test results on a checkpoint, run:

cd FCN
./visualise.sh

UNet

For ablations and visualisations, see the report.

Visualisations

Each image has three columns, representing input image, ground truth, followed by prediction (each image is a collage of these three in that order).

Variant	Prediction 1	Prediction 2
Vanilla
Residual
No-Skip
Gated Attention

To train and test the model, run:

cd UNet
python -m src.main <options>

The options include the data directory, batch size, model variant, the number of epochs, and the learning rate. See the main.py file for more details.

To visualise some test results on a checkpoint, run:

cd UNet
./visualise.sh

Setup

Download the data from the following link.

The environment can be set up with the conda env file:

cd docs
conda env create -f env.yml

or

pip install -r requirements.txt

Alternatively, install the dependencies as in the conda history.

Scripts

Helper utilities are available to split the train data into train and validation for the FCN, and for the UNet, as well as to perform some exploratory data analysis for the FCN.