OCN: Learning Object-centric Representations for Unsupervised Multi-object Segmentation

ICLR 2025 Conference Submission472 Authors

Keywords: unsupervised learning, object segmentation, objectness representation

TL;DR: We propose a two stage pipeline for unsupervised multi-object segmentation on single images by learning and reasoning with three level object-centric representations.

Abstract: We study the challenging problem of unsupervised multi-object segmentation on single images. By relying on an image reconstruction objective to learn objectness or leveraging pretrained image features to group similar pixels as objects, most existing methods can either segment simple synthetic objects or discover a rather limited number of real-world objects. In this paper, we introduce OCN, a new two stage pipeline to discover many complex objects on real-world images. The key to our approach is to explicitly learn our carefully defined three level object-centric representations in the first stage. After that, our multi-object reasoning module directly leverages the learned object priors to discover multiple objects in the second stage. Notably, such a reasoning module is completely network-free and does not need any human labels to train. Extensive experiments show that our OCN clearly surpasses all existing unsupervised methods by a large margin on 7 real-world benchmark datasets including the particularly challenging COCO dataset, achieving the state-of-the-art object segmentation results. Most notably, our method demonstrates superior results on extremely crowded images where all baselines collapse.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 472