OmniContrast: Vision-Language-Interleaved Contrast from Pixels All at once
Keywords: vision-language contrastive learning
Abstract: In this work, we present OmniContrast, a unified contrastive learning model tailored for vision, language, and vision-language-interleaved understanding within multi-modal web documents. Unlike traditional image-caption data with clear vision-language correspondence, we explore a new contrastive fashion on maximizing the similarity between consecutive snippets sampled from image-text interleaved web documents. Moreover, to enable CLIP to handle long-form text and image-text interleaved content from web documents, OmniContrast unifies all modalities into pixel space, where text is rendered visually. This unification simplifies the processing and representation of diverse multi-modal inputs, enabling a single vision model to process any modality. To evaluate the omni-modality understanding of OmniContrast, we design three consecutive information retrieval benchmarks AnyCIR, SeqCIR, and CSR. Extensive experimental results demonstrate that OmniContrast achieves superior or competitive omni-modality understanding performance to existing standard CLIP models trained on image-text pairs. This highlights the potential of multi-modal web documents as a rich and valuable resource for advancing vision-language learning.
Primary Area: applications to computer vision, audio, language, and other modalities
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: 2906
Loading