SINDER: Repairing the Singular Defects of DINOv2
This addresses a specific defect in vision models for researchers and practitioners, offering an incremental improvement by fine-tuning with minimal data.
The paper tackled artifacts in patch tokens of Vision Transformer models like DINOv2, identifying their origin in the pre-trained network's leading left singular vector, and proposed a fine-tuning smooth regularization that improves performance on tasks such as unsupervised segmentation and classification without full re-training.
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.