Self-supervised visual foundation models produce powerful embeddings that achieve remarkable performance on a wide range of downstream tasks. However, unlike vision-language models such as CLIP, self-supervised visual features are not readily aligned with language, hindering their adoption in open-vocabulary tasks. Our method, named dino.txt, unlocks this new ability for DINOv2, a widely used self-supervised visual encoder. We build upon the LiT training strategy, which trains a text encoder to align with a frozen vision model but leads to unsatisfactory results on dense tasks. We propose several key ingredients to improve performance on both global and dense tasks, such as concatenating the [CLS] token with the patch average to train the alignment and curating data using both text and image modalities. With these, we successfully train a CLIP-like model with only a fraction of the computational cost compared to CLIP while achieving state-of-the-art results in zero-shot classification and open-vocabulary semantic segmentation.
@article{arxiv.2412.16334,
title = {DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignment},
author = {Cijo Jose and Théo Moutakanni and Dahyun Kang and Federico Baldassarre and Timothée Darcet and Hu Xu and Daniel Li and Marc Szafraniec and Michaël Ramamonjisoa and Maxime Oquab and Oriane Siméoni and Huy V. Vo and Patrick Labatut and Piotr Bojanowski},
journal= {arXiv preprint arXiv:2412.16334},
year = {2024}
}