This repository contains supplementary material that belongs to the research I conducted for my master's thesis. A link to the final version of my thesis will be posted here soon. Among other content, the code and data, which are needed to reproduce the results, are included.
This repository allows for fast reproduction of the occlusion experiments. No downloads and installations are required because the code can be ran in the cloud for free, using Google Colaboratory.
The only thing you have to do to get started is:
- Navigate to the "colab_notebooks" folder.
- Open one of the notebooks by clicking on it.
- Click the blue "Open in Colab" button at the beginning of the notebook.
- Follow the instructions in the notebook (or just run all cells of the notebook by pressing ctrl+F9).
- Recommended: play around with the parameters as explained in the notebooks.
Additionally, notebooks are provided that can be used to reproduce the model training and tuning processes. These notebooks are also helpful for training and evaluating models with new, improved, and/or different configurations. These configurations have to be specified in a ".config" file.
In the "main_content/assortment of heatmaps" folder, a number of heatmaps can be found, many of which are not included in my thesis.
The work here relies for a great deal on the TensorFlow object detection API (https://github.com/tensorflow/models/tree/master/research/object_detection). So, thanks a lot to the developers and maintainers of that repository.