The training and evaluation of this network was done in Google Colaboratory, so it's not straightforward to run the notebooks because the paths are different.
The results are explained in the report and can be visualised by looking at the predicted outputs of the network.
The repo is organised as follows:
- colab_notebooks directory contains:
- <model>_pretrained.ipynb notebooks for building and training each one of the chosen backbones, using both our own implementation and the Segmentation Models Library.
- models_evaluator notebook for evaluating each model both by its metrics and looking at the predictions.
- dataset_filtering directory contains:
- filtered_cats() and most_annotated() functions used to filter the MS-COCO dataset by categories.
- DataGeneration class dedicated to generate the input and output samples for the net.
- model directory contains test files regarding the construction, training and loading of Keras models.
- tests directory contains all kind of test files for the modules made, and more.
- The dataset should be added at a data/ folder from the root directory
Authors: