Emerging Properties in Self-Supervised Vision Transformers

In this paper, we question if self-supervised learning provides new properties to Vision Transformer (ViT) that stand out compared to convolutional networks (convnets). Beyond the fact that adapting self-supervised methods to this architecture works particularly well, we make the following observations: first, self-supervised ViT features contain explicit information about the semantic segmentation of an image, which does not emerge as clearly with supervised ViTs, nor with convnets... Second, these features are also excellent k-NN classifiers, reaching 78.3% top-1 on ImageNet with a small ViT. Our study also underlines the importance of momentum encoder, multi-crop training, and the use of small patches with ViTs. We implement our findings into a simple self-supervised method, called DINO, which we interpret as a form of self-distillation with no labels. We show the synergy between DINO and ViTs by achieving 80.1% top-1 on ImageNet in linear evaluation with ViT-Base. read more

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Copy Detection Copydays strong subset DINO (ViT-B/8) mAP 85.5 # 1
Video Object Detection DAVIS 2017 DINO (ViT-B/8, ImageNet retrain) J&F 71.4 # 1
Self-Supervised Image Classification ImageNet DINO (ViT-B/16) Top 1 Accuracy 78.2% # 15
Number of Params 85M # 27
Top 1 Accuracy (kNN, k=20) 76.1 # 8
Self-Supervised Image Classification ImageNet DINO (ViT-B/8) Top 1 Accuracy 80.1% # 6
Number of Params 85M # 27
Top 1 Accuracy (kNN, k=20) 77.4 # 5
Self-Supervised Image Classification ImageNet DINO (ViT-S/16) Top 1 Accuracy 77.0% # 22
Number of Params 21M # 53
Top 1 Accuracy (kNN, k=20) 74.5 # 10
Self-Supervised Image Classification ImageNet DINO (ResNet-50) Top 1 Accuracy 75.3% # 31
Number of Params 24M # 38
Top 1 Accuracy (kNN, k=20) 67.5 # 13
Self-Supervised Image Classification ImageNet DINO (ViT-S/8) Top 1 Accuracy 79.7% # 8
Number of Params 21M # 53
Top 1 Accuracy (kNN, k=20) 78.3 # 3
Self-Supervised Image Classification ImageNet (finetuned) DINO (ViT-B/16) Number of Params 85M # 4
Top 1 Accuracy 82.8 # 5

Methods