SINDER: Repairing the Singular Defects of DINOv2

TL;DR

This paper investigates artifacts in DINOv2 Vision Transformer models, identifies their origin in the leading singular vector of weights, and proposes a fine-tuning regularization method to repair these defects efficiently.

Contribution

The paper uncovers the root cause of patch token artifacts in Vision Transformers and introduces a novel regularization technique for effective model repair without full re-training.

Findings

Artifacts originate from the leading singular vector of weights.

Proposed regularization improves downstream task performance.

Method avoids full re-training, saving computational resources.

Abstract

Vision Transformer models trained on large-scale datasets, although effective, often exhibit artifacts in the patch token they extract. While such defects can be alleviated by re-training the entire model with additional classification tokens, the underlying reasons for the presence of these tokens remain unclear. In this paper, we conduct a thorough investigation of this phenomenon, combining theoretical analysis with empirical observations. Our findings reveal that these artifacts originate from the pre-trained network itself, specifically stemming from the leading left singular vector of the network's weights. Furthermore, to mitigate these defects, we propose a novel fine-tuning smooth regularization that rectifies structural deficiencies using only a small dataset, thereby avoiding the need for complete re-training. We validate our method on various downstream tasks, including unsupervised segmentation, classification, supervised segmentation, and depth estimation, demonstrating its effectiveness in improving model performance. Codes and checkpoints are available at https://github.com/haoqiwang/sinder.