Analyzing Local Representations of Self-supervised Vision Transformers
This work addresses the problem of understanding and improving local representations in self-supervised vision models for researchers and practitioners in computer vision, but it is incremental as it builds on existing methods.
The paper analyzes local representations of self-supervised Vision Transformers, finding that contrastive learning methods like DINO produce more universal patch representations for downstream tasks without fine-tuning, while masked image modeling methods like MAE have high-variance features that harm performance, and it shows that DINOv2 underperforms DINO in some retrieval tasks despite more data.
In this paper, we present a comparative analysis of various self-supervised Vision Transformers (ViTs), focusing on their local representative power. Inspired by large language models, we examine the abilities of ViTs to perform various computer vision tasks with little to no fine-tuning. We design evaluation framework to analyze the quality of local, i.e.\ patch-level, representations in the context of few-shot semantic segmentation, instance identification, object retrieval and tracking. We discover that contrastive learning based methods like DINO produce more universal patch representations that can be immediately applied for downstream tasks with no parameter tuning, compared to masked image modeling. The embeddings learned using the latter approach, e.g. in masked autoencoders, have high variance features that harm distance-based algorithms, such as k-NN, and do not contain useful information for most downstream tasks. Furthermore, we demonstrate that removing these high-variance features enhances k-NN for MAE, as well as for its recent extension Scale-MAE. Finally, we find an object instance retrieval setting where DINOv2, a model pretrained on two orders of magnitude more data, falls short of its less compute intensive counterpart DINO.