DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignment

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Jose, Cijo, Moutakanni, Théo, Kang, Dahyun, Baldassarre, Federico, Darcet, Timothée, Xu, Hu, Li, Daniel, Szafraniec, Marc, Ramamonjisoa, Michaël, Oquab, Maxime, Siméoni, Oriane, Vo, Huy V., Labatut, Patrick, Bojanowski, Piotr
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913622600450048
author Jose, Cijo
Moutakanni, Théo
Kang, Dahyun
Baldassarre, Federico
Darcet, Timothée
Xu, Hu
Li, Daniel
Szafraniec, Marc
Ramamonjisoa, Michaël
Oquab, Maxime
Siméoni, Oriane
Vo, Huy V.
Labatut, Patrick
Bojanowski, Piotr
author_facet Jose, Cijo
Moutakanni, Théo
Kang, Dahyun
Baldassarre, Federico
Darcet, Timothée
Xu, Hu
Li, Daniel
Szafraniec, Marc
Ramamonjisoa, Michaël
Oquab, Maxime
Siméoni, Oriane
Vo, Huy V.
Labatut, Patrick
Bojanowski, Piotr
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2412_16334
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignment
Jose, Cijo
Moutakanni, Théo
Kang, Dahyun
Baldassarre, Federico
Darcet, Timothée
Xu, Hu
Li, Daniel
Szafraniec, Marc
Ramamonjisoa, Michaël
Oquab, Maxime
Siméoni, Oriane
Vo, Huy V.
Labatut, Patrick
Bojanowski, Piotr
Computer Vision and Pattern Recognition
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.
title DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignment
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2412.16334