# DINOv2: Learning Robust Visual Features without Supervision

**Authors:** Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve  
**arXiv:** [2304.07193](https://arxiv.org/abs/2304.07193)  
**Date:** 2023-04-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.07193v1/

## Key points

- Natural language processing for model pretraining has led to a breakthrough in computer vision
- All-purpose visual features can be produced without requiring finetuning
- Self-supervised methods have shown great potential in producing such features with curated data from diverse sources
- Researchers have combined different techniques to accelerate and stabilize training at scale
- An automatic pipeline was proposed to build a dedicated, diverse, and curated image dataset instead of using uncurated data
- A ViT model with 1B parameters was trained and distilled into smaller models that surpassed the best available all-purpose features on most benchmarks
- The findings have significant implications for simplifying the use of images in any system and advancing computer vision research.

## Summary

The field of computer vision has seen a significant breakthrough with the development of natural language processing for model pretraining on large quantities of data. This has opened up new possibilities for creating foundation models in computer vision that can produce all-purpose visual features, which work across image distributions and tasks without requiring finetuning. In this regard, existing pretraining methods, particularly self-supervised methods, have shown great potential in producing such features if trained on enough curated data from diverse sources. To further scale the pretraining process in terms of data and model size, researchers have revisited existing approaches and combined different techniques to accelerate and stabilize training at scale. One of the key contributions is an automatic pipeline proposed to build a dedicated, diverse, and curated image dataset instead of using uncurated data as typically done in the self-supervised literature. Additionally, a ViT model with 1B parameters was trained and distilled into a series of smaller models that surpassed the best available all-purpose features, OpenCLIP (Ilharco et al., 2021), on most benchmarks at both image and pixel levels. The team behind this research includes Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa , Alaaeldin El-Nouby , Mahmoud Assran , Nicolas Ballas , Wojciech Galuba , Russell Howes , Po-Yao Huang , Shang-Wen Li , Ishan Misra , Michael Rabbat , Vasu Sharma , Gabriel Synnaeve , Hu Xu , Hervé Jegou , Julien Mairal Patrick Labatut Armand Joulin Piotr Bojanowski. Overall, this work demonstrates that it is possible to learn robust visual features without supervision by leveraging pretraining methods on curated datasets. The findings have significant implications for simplifying the use of images in any system and advancing computer vision research.