Research map / Architectures and efficiency / Pruning and efficient tuning
SINDER: Repairing the Singular Defects of DINOv2
Haoqi Wang, Tong Zhang, Mathieu Salzmann
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| repair singular defects with SINDER fine-tuning | DINOv2 giant with STEGO · Cityscapes unsupervised | mIoU: 19.38 -> 21.77 | Improved |
| repair singular defects with SINDER fine-tuning | DINOv2 giant with CAUSE · VOC2012 without CRF | mIoU: 55.8 -> 62.9 | Improved |
| repair singular defects with SINDER fine-tuning | DINOv2 giant · ImageNet-1K linear probe | top-1 accuracy: 86.53 -> 86.29 | No clear effect |
| repair singular defects with SINDER fine-tuning | DINOv2 giant · ADE20k linear segmentation | mIoU: 48.83 -> 51.11 | Improved |
| SINDER versus register-token retraining | DINOv2 giant · NYUd depth DPT | RMSE: 0.234 -> 0.249 | Hurt |
| repair singular defects with SINDER fine-tuning | DINOv2 giant · NYUd depth Linear 1 | RMSE: 0.370 -> 0.337 | Improved |
| also fine-tune biases along with singular values (ablation) | SINDER · ImageNet KNN and ADE20k | ADE20k mIoU: 51.11 -> 13.77 | Hurt |
| limit learnable layers to 10 per iteration instead of all layers (ablation) | SINDER · ADE20k | mIoU: 45.53 -> 51.11 | Improved |
| limit learnable layers to 5 instead of 10 (ablation) | SINDER · ADE20k | mIoU: 51.11 -> 50.61 | Hurt |
| raise skip threshold sigma from 3 to 5 (ablation) | SINDER · ADE20k | mIoU: 51.11 -> 50.71 | Hurt |
| raise mask threshold mu from 4 to 4.5 (ablation) | SINDER · ADE20k | mIoU: 51.11 -> 50.50 | Hurt |
Closest papers by meaning
- Simplifying DINO via Coding Rate Regularization (ICML 2025 · similarity 0.85)
- Bootstrapping SparseFormers from Vision Foundation Models (CVPR 2024 · similarity 0.84)
- Vision Transformers with Self-Distilled Registers (NeurIPS 2025 · similarity 0.84)
- Learning Where to Edit Vision Transformers (NeurIPS 2024 · similarity 0.84)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.84)
- Low-Rank Rescaled Vision Transformer Fine-Tuning: A Residual Design Approach (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
