IBB Assignment 3

This repository contains source code of my solution of Assignment 3 for Image Based Biometry course at University of Ljubljana.

Report with IMRAD structure is available as release asset.

Requirements

Python 3.8.2 was used with the following packages installed:

matplotlib==3.3.2
numpy==1.18.5
pandas==1.1.3
tensorflow==2.3.1

Additionally, folder data must contain AWE dataset unzipped so that e.g. data/001/01.png is a valid path.

Loading and training

Script train.py can be run as-is. It will download EfficientNet-B0 weights and load saved weights of our CNN model trained without image augmentations. To switch to model with image augmentations, change parameters near top of the file to contain:

EXP_ID = "model-b"
AUGMENTATIONS = True

To enable model training, change parameters to:

TRAIN = True

Evaluation

Script evaluate.py plots figures (to folder figures) and prints performance metrics to console. It uses state of models provided in folder out. This state can be recomputed by executing script train.py (once with augmentations and once without them) as described in previous section.

Report

Source code of LaTeX report is contained in report/jj1712.tex. Before compiling it, make sure you have generated figures as described in Evaluation.