Javascript must be enabled for the correct page display

Exploring Self-Supervised Learning with Geometric Transformations

Vries, Tanja de (2023) Exploring Self-Supervised Learning with Geometric Transformations. Master's Thesis / Essay, Computing Science.


Download (2MB) | Preview
[img] Text
Restricted to Registered users only

Download (134kB)


We made a self-supervised model with a pretext-task based on geometric transformations. Our model is based on RotNet, a model that predicts image rotations by Gidaris et al. [GSK18]. We implemented our model in TensorFlow, and were able to reproduce their results for rotation. The rotation model cannot be used for all datasets, for example datasets with a lot of top-down images or images of round objects. Therefore, we modified our network to predict the scale of an image. In order for our scale model to work, the used dataset needs to have a defined scale, that is, all images are taken from the same distance. We see the patterns that show the pretext task learns some useful features. However, we also see that for most experiments the gap with supervised is bigger for our scale model than for our rotation model. Most promising is that our scale model closes the gap with supervised learning during our experiment with a low amount of labelled data.

Item Type: Thesis (Master's Thesis / Essay)
Supervisor name: Biehl, M.
Degree programme: Computing Science
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 31 Jul 2023 08:43
Last Modified: 31 Jul 2023 08:43

Actions (login required)

View Item View Item