Research map / Label-efficient and robust learning / Self-supervised and contrastive
Ditch the Denoiser: Emergence of Noise Robustness in Self-Supervised Learning from Data Curriculum
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a denoiser-free noise-robust SSL pipeline (noise curriculum plus denoised-teacher regularization) that pretrains DINOv2 on denoised-then-noisy data so the denoiser can be discarded at inference.
Problem
Self-supervised learning degrades substantially when pretrained on noisy images, and the standard fix of a denoiser preprocessing step adds latency, fine-tuning overhead, and deployment complexity.
Approach
Train an SSL denoiser (Neighbor2Neighbor) on the noisy data, pretrain DINOv2 on denoised images first and then restart training on the noisy images as an entropy-increasing curriculum, and optionally anchor noisy-branch embeddings to a frozen denoised-trained teacher, so the final backbone works on raw noisy inputs without a denoiser.
Key results
- On ImageNet-1k with ViT-B under Gaussian sigma=255 (0.72 dB), DINOv2 w/ NCT improves linear probing accuracy from 51.0% to 55.8% (+4.8) over DINOv2, versus 57.2% for N2N + DINOv2.
- On ImageNet-100 ViT-S with Gaussian sigma=255 at 200 epochs, DINOv2 w/ NC reaches 47.9% vs 35.8% for DINOv2, and DINOv2 w/ NCT reaches 54.6%.
- On the clean ImageNet validation set, DINOv2 w/ NC or NCT outperforms N2N + DINOv2 in most cases, e.g., 65.6 vs 64.1 for NCT at Gaussian-255 on ImageNet-1k.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add denoised-to-noisy curriculum | DINOv2 ViT-B/16 · ImageNet-1k, Gaussian-255 | accuracy: 51.0 -> 53.5 | Helped |
| add denoised-teacher regularization to the curriculum | DINOv2 ViT-B/16 · ImageNet-1k, Gaussian-255 | accuracy: 53.5 -> 55.8 | Helped |
| add denoised-to-noisy curriculum | DINOv2 ViT-B/16 · ImageNet-1k, Gaussian-100 | accuracy: 70.7 -> 70.9 | Helped |
| add denoised-teacher regularization to the curriculum | DINOv2 ViT-B/16 · ImageNet-1k, Gaussian-100 | accuracy: 70.9 -> 72.1 | Helped |
| add denoised-to-noisy curriculum | DINOv2 ViT-S/16 · ImageNet-100, Gaussian-50 | accuracy: 65.0 -> 72.8 | Helped |
| add denoised-to-noisy curriculum | DINOv2 ViT-S/16 · ImageNet-100, Gaussian-100 | accuracy: 55.4 -> 68.1 | Helped |
| add denoised-teacher regularization to the curriculum | DINOv2 ViT-S/16 · ImageNet-100, Gaussian-255 | accuracy: 47.9 -> 54.6 | Helped |
| increase training duration from 200 to 1000 epochs | DINOv2 · ImageNet-100, Gaussian-255 | accuracy: 35.8 -> 56.8 | Helped |
| replace noisy-only training with denoised pretraining followed by noisy training (ablation) | DINOv2 ViT-S/16 · ImageNet-100, Gaussian-100 | accuracy | Helped |
| restart training dynamics for the noisy training stage (ablation) | DINOv2 ViT-S/16 · ImageNet-100, Gaussian-100 | accuracy | Helped |
| use a separately trained, unaligned frozen teacher instead of the aligned denoised teacher (ablation) | DINOv2 w/ NCT · ImageNet-100, Gaussian-255 | accuracy: +1.1% | Hurt |
| replace denoised-pretrained trainable backbones with randomly initialized backbones (ablation) | DINOv2 w/ NCT · ImageNet-100, Gaussian-255 | accuracy | Hurt |
Surprising findings
- The denoiser-free DINOv2 w/ NC/NCT often beats the two-stage N2N + DINOv2 pipeline on clean validation accuracy.
- Even a denoiser trained for only 1 epoch gives a large gain (55.4 to 66.1) when used for the curriculum.
Limitations
- Evaluated only on synthetic noise (Gaussian, shot, speckle) on ImageNet variants, not real-world noisy datasets.
- Relies on the self-supervised denoiser being reasonably good; the curriculum design (restart epoch, regularization strength) requires tuning.
- Regularization strength lambda and restart epoch are selected via sweeps, and the sweep was not uniform across settings.
Experimental setup
Datasets: ImageNet-100 ImageNet-1k Oxford Paris MNIST
Benchmarks: ImageNet-100 ImageNet-1k Oxford Paris
Metrics: linear probing top-1 accuracy mean average precision (mAP)
Baselines: N2N + DINOv2 DINOv2 SimCLR DINOv2 w/ NC MoCo v3 SimSiam iBOT DINO
Builds on: DINOv2 Neighbor2Neighbor curriculum learning
Closest papers by meaning
- Convexity-Aware Noise Calibration: A Self-Supervised Framework for Noise-Level-Unknown Image Denoising (CVPR 2026 · similarity 0.90)
- UNSURE: self-supervised learning with Unknown Noise level and Stein's Unbiased Risk Estimate (ICLR 2025 · similarity 0.89)
- Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising (CVPR 2025 · similarity 0.87)
- Exploring Efficient Asymmetric Blind-Spots for Self-Supervised Denoising in Real-World Scenarios (CVPR 2024 · similarity 0.86)
- Robust Representation Consistency Model via Contrastive Denoising (ICLR 2025 · similarity 0.86)
- Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.86)
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.