Unsupervised Whole Object Discovery by Contextual Grouping with Repulsion

ICLR 2025 Conference Submission1347 Authors

Keywords: Unsupervised Object Discovery, Unsupervised Whole Object Segmentation, Co-Segmentation, Normalized Cut, Attraction and Repulsion

Abstract: It is challenging to discover and segment whole objects from unlabeled images, as features unsupervisedly learned on images tend to focus on distinctive appearances (e.g., the face rather than the torso), and grouping by feature similarity could reveal only these representative parts, not the whole objects (e.g., the entire human body). Our key insight is that, an object of distinctive parts pops out as a whole, due not only to how similar they are to each other, but also to it how different they are from their contexts within an image or across related images. The latter could be crucial for binding different parts into a coherent whole without preconception of objects. We formulate our idea for unsupervised object segmentation in a spectral graph partitioning framework, where nodes are patches and edges are grouping cues between patches, measured by feature similarity for attraction, and by feature dissimilarity for repulsion. We seek the graph cuts that maximize within-group attraction and figure-ground repulsion while minimizing figure/ground attraction and within-group repulsion. Our simple method consistently outperforms the state-of-the-art on unsupervised object discovery, figure/ground saliency detection, and unsupervised video object segmentation benchmarks. In particular, it excels at discovering whole objects instead of salient parts.

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: 1347