DINOv2: Learning Robust Visual Features without Supervision

TLDR

Recent breakthroughs in NLP pretraining have paved the way for foundation models in computer vision that produce all‑purpose visual features usable across tasks without finetuning. The authors aim to scale self‑supervised pretraining by combining techniques and building a curated dataset to accelerate and stabilize training at scale. They train a 1‑billion‑parameter ViT on a curated, diverse dataset and distill it into smaller models, using an automated pipeline to assemble the dataset and scale training procedures. The resulting models outperform the best available all‑purpose features, such as OpenCLIP, on most image‑ and pixel‑level benchmarks, demonstrating that large‑scale, curated self‑supervised pretraining yields robust visual features.

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.