Multi-view Object-Centric Learning with Identifiable Representations
Keywords: multi-view object-centric learning, identifiability, probabilistic-slot-attention
TL;DR: We propose a method to learn identifiable multi-view object-centric representations up to a equivalence relation, resolving spatial ambiguities.
Abstract: Modular object-centric representations are key to unlocking human-like reasoning capabilities. However, addressing challenges such as object occlusions to obtain meaningful object-level representations presents both theoretical and practical difficulties. We introduce a novel multi-view probabilistic approach that aggregates view-specific slots to capture *invariant content* information while simultaneously learning disentangled global *viewpoint-level* information. Our model resolves spatial ambiguities and provides theoretical guarantees for learning identifiable representations, setting it apart from prior work focusing on single-view settings and lacking theoretical foundations. Along with our identifiability analysis, we provide extensive empirical validation with promising results on both benchmark and proposed large-scale datasets carefully designed to evaluate multi-view methods.
Primary Area: generative models
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.
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: 2085
Loading