DINOv2: Learning Robust Visual Features without Supervision

TL;DR

This paper introduces DINOv2, a self-supervised vision model trained on a large, curated dataset, producing robust, all-purpose visual features that outperform existing models across various tasks without fine-tuning.

Contribution

It presents a scalable self-supervised training approach with a curated dataset and a large ViT model, achieving state-of-the-art all-purpose visual features.

Findings

Large-scale self-supervised training improves feature robustness.

DINOv2 surpasses previous models like OpenCLIP on multiple benchmarks.

Curated datasets enhance training stability and model performance.

Abstract

The recent breakthroughs in natural language processing for model pretraining on large quantities of data have opened the way for similar foundation models in computer vision. These models could greatly simplify the use of images in any system by producing all-purpose visual features, i.e., features that work across image distributions and tasks without finetuning. This work shows that existing pretraining methods, especially self-supervised methods, can produce such features if trained on enough curated data from diverse sources. We revisit existing approaches and combine different techniques to scale our pretraining in terms of data and model size. Most of the technical contributions aim at accelerating and stabilizing the training at scale. In terms of data, we propose an automatic pipeline to build a dedicated, diverse, and curated image dataset instead of uncurated data, as typically done in the self-supervised literature. In terms of models, we train a ViT model (Dosovitskiy et al., 2020) with 1B parameters and distill it into a series of smaller models that surpass the best available all-purpose features, OpenCLIP (Ilharco et al., 2021) on most of the benchmarks at image and pixel levels.