Research map / Label-efficient and robust learning / Noisy labels
Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data
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 iMRF, an MRF-based label-correction method with agreement-based adaptive thresholds and a noise-aware balanced cross-entropy loss, for learning from class-imbalanced data containing both closed-set and open-set noise.
Problem
Learning robust visual recognition from open-world noisy data (both closed-set and open-set label noise) when the label distribution is long-tailed and class-imbalanced, where prior methods assume balanced classes.
Approach
An iterative MRF over a KNN graph of all training samples, inferred batch-wise per mini-batch, corrects closed-set noisy labels; a tripartite agreement (given label, current and previous inference) sets adaptive thresholds to select clean samples and reject open-set noise; a noise-aware balanced softmax loss uses label frequencies from the selected clean and corrected sets to remove class-imbalanc
Key results
- On CIFAR-100 (imbalance factor 10, 50% symmetric noise, 20%/40% TinyImageNet open-set), the method reaches about 56% top-1 versus 54% for NGC in the 20% case and 56.5% vs 51.8% at 40%.
- On WebVision-50 the method reports 80.80% top-1 and 93.00% top-5, above NGC's 78.92% and 92.24%.
- Ablation on CIFAR-100 shows removing the adaptive thresholding strategy causes the largest performance drop among components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use iMRF label correction instead of network predictions as posterior (ablation) | Ours · CIFAR-100, LIOND, TinyImageNet open-set ρ=40% | top-1 accuracy (%): 53.56 -> 56.54 | Helped |
| Use adaptive agreement-based thresholding instead of fixed thresholds (ablation) | Ours · CIFAR-100, LIOND, TinyImageNet open-set ρ=40% | top-1 accuracy (%): 49.92 -> 56.54 | Helped |
| Use noisy-aware balanced cross-entropy instead of conventional cross-entropy (ablation) | Ours · CIFAR-100, LIOND, TinyImageNet open-set ρ=40% | top-1 accuracy (%): 54.34 -> 56.54 | Helped |
| Add mixup loss (ablation) | Ours · CIFAR-100, LIOND, TinyImageNet open-set ρ=40% | top-1 accuracy (%): 51.14 -> 56.54 | Helped |
| Add supervised contrastive loss (ablation) | Ours · CIFAR-100, LIOND, TinyImageNet open-set ρ=40% | top-1 accuracy (%): 55.66 -> 56.54 | Helped |
| Add instance-wise contrastive loss (ablation) | Ours · CIFAR-100, LIOND, TinyImageNet open-set ρ=40% | top-1 accuracy (%): 55.41 -> 56.54 | Helped |
| Add cross-view consistency loss (ablation) | Ours · CIFAR-100, LIOND, TinyImageNet open-set ρ=40% | top-1 accuracy (%): 54.78 -> 56.54 | Helped |
| Use the proposed method instead of NGC | CIFAR-10 LIOND · CIFAR-10, μ=10, κ=50%, with open-set noise | top-1 accuracy (%): 86.15±0.28 -> 86.66±0.38 | Helped |
| Use the proposed method instead of NGC | CIFAR-100 LIOND · CIFAR-100, μ=10, κ=50%, with open-set noise | top-1 accuracy (%): 53.89±0.88 -> 56.03±0.68 | Helped |
| Use the proposed method instead of NGC | Ours on WebVision-50 · WebVision-50 validation | top-1 accuracy: 78.92 -> 80.80 | Helped |
| Use the proposed method instead of NGC | Ours on ILSVRC12 · ILSVRC12 validation | top-1 accuracy: 75.00 -> 77.24 | Helped |
| Increase K to 100 (ablation) | iMRF · CIFAR-100, LIOND | top-1 accuracy (%) | Hurt |
Surprising findings
- Larger K (e.g., K=100) harms performance because tail classes with very few samples cannot form valid neighborhoods.
Limitations
- Relies on fixed hyperparameters (alpha, beta, gamma, K) and a warm-up stage, with performance sensitive to large K for tail classes.
- Evaluation on CIFAR uses simulated noise and open-set injection rather than fully real-world noise.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 TinyImageNet Places-365 WebVision-50
Benchmarks: CIFAR-10 CIFAR-100 WebVision-50 ILSVRC12
Metrics: top-1 accuracy top-5 accuracy
Baselines: Jo-SRC ProtoMix NGC INCV DivideMix ELR+
Builds on: ProtoMix NGC Jo-SRC balanced softmax mixup supervised contrastive learning
Closest papers by meaning
- Domain-Conditioned Class Imbalance: Why Global Class Balance Fails Across Domains (NeurIPS 2026 · similarity 0.89)
- Rethinking Class-Balanced Methods for Long-Tailed Visual Recognition From a Domain Adaptation Perspective (CVPR 2020 · similarity 0.89)
- Large-Scale Long-Tailed Recognition in an Open World (CVPR 2019 · similarity 0.89)
- Rethinking the Value of Labels for Improving Class-Imbalanced Learning (NeurIPS 2020 · similarity 0.88)
- CUDA: Curriculum of Data Augmentation for Long-tailed Recognition (ICLR 2023 · similarity 0.88)
- Learning with Open-world Noisy Data via Class-independent Margin in Dual Representation Space (AAAI 2025 · similarity 0.87)
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.