Open-Vocabulary Object Detection for Incomparable Spaces
Keywords: Multimodal learning, object detection
Abstract: In open-vocabulary object detection (OVDet), specifying the object of interest at inference time opens up powerful possibilities, allowing users to define new categories without retraining the model. These objects can be identified through text descriptions, image examples, or a combination of both. However, visual and textual data, while complementary, encode different data types, making direct comparison or alignment challenging. Naive fusion approaches often lead to misaligned predictions, particularly when one modality is ambiguous or incomplete. In this work, we propose an approach for OVDet that aligns relational structures across these incomparable spaces, ensuring optimal correspondence between visual and textual inputs. This shift from feature fusion to relational alignment bridges the gap between these spaces, enabling robust detection even when input from one modality is weak. Our evaluation on the challenging datasets demonstrates that our model sets a new benchmark in detecting rare objects, outperforming existing OVDet models. Additionally, we show that our multi-modal classifiers outperform single-modality models and even surpass fully-supervised detectors.
Primary Area: learning theory
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: 10775
Loading