Unsupervised Object Segmentation with Generative Models

Bielski, Adam Jakub (2024). Unsupervised Object Segmentation with Generative Models (Submitted). (Dissertation)

Text PhD_Thesis_Adam_Bielski.pdf - Other Restricted to registered users only Available under License BORIS Standard License. Download (48MB)

Advances in computer vision have transformed how we interact with technology, driven
by significant breakthroughs in scalable deep learning and the availability of large
datasets. These technologies now play a crucial role in various applications, from
improving user experience through applications like organizing digital photo libraries,
to advancing medical diagnostics and treatments. Despite these valuable applications,
the creation of annotated datasets remains a significant bottleneck. It is not only costly
and labor-intensive but also prone to inaccuracies and human biases. Moreover, it often
requires specialized knowledge or careful handling of sensitive information. Among the
tasks in computer vision, image segmentation particularly highlights these challenges,
with its need for precise pixel-level annotations. This context underscores the need for
unsupervised approaches in computer vision, which can leverage the large volumes of
unlabeled images produced every day.
This thesis introduces several novel methods for learning fully unsupervised object
segmentation models using only collections of images. Unlike much prior work, our
approaches are effective on complex real-world images and do not rely on any form
of annotations, including pre-trained supervised networks, bounding boxes, or class
labels. We identify and leverage intrinsic properties of objects – most notably, the
cohesive movement of object parts – as powerful signals for driving unsupervised
object segmentation. Utilizing innovative generative adversarial models, we employ this
principle to either generate segmented objects or directly segment them in a manner
that allows for realistic movement within scenes. Our work demonstrates how such
generated data can train a segmentation model that effectively generalizes to realworld
images. Furthermore, we introduce a method that, in conjunction with recent
advances in self-supervised learning, achieves state-of-the-art results in unsupervised
object segmentation. Our methods rely on the effectiveness of Generative Adversarial
Networks, which are known to be challenging to train and exhibit mode collapse. We
propose a new, more principled GAN loss, whose gradients encourage the generator
model to explore missing modes in its distribution, addressing these limitations and
enhancing the robustness of generative models.

Interest & Impact

Downloads

1 since deposited on 27 May 2024
1 in the past 12 months

Citations

5 Citations in Web of Science ®
6 Citations in Scopus

Search

in Google Scholar™

Services

Actions (login required)

Edit item

Actions (login required)

Edit item