Research map / Label-efficient and robust learning / Noisy labels
Re-Labeling ImageNet: From Single to Multi-Labels, From Global to Localized Labels
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train with localized multi-labels instead of original single labels | ResNet-50 · ImageNet validation | top-1 accuracy: 77.5 -> 78.9 (+1.4 pp) | Helped |
| Train with localized multi-labels instead of original single labels | ResNet-50 · ReaL | multi-label accuracy: 83.6 -> 85.0 | Helped |
| Train with localized multi-labels instead of original single labels | ResNet-50 · Shankar multi-label evaluation | multi-label accuracy: 85.3 -> 86.1 | Helped |
| Apply label smoothing instead of original labels | ResNet-50 · Shankar multi-label evaluation | multi-label accuracy: 85.3 -> 84.7 | Hurt |
| Use localized single-label targets rather than localized multi-label targets (ablation) | ReLabel on ResNet-50 · ImageNet validation | top-1 accuracy: 78.9 -> 78.4 (-0.5) | Hurt |
| Use global multi-label targets rather than localized targets (ablation) | ReLabel on ResNet-50 · ImageNet validation | top-1 accuracy: 78.9 -> 78.5 (-0.4) | Hurt |
| Remove both multi-label targets and localization (ablation) | ReLabel on ResNet-50 · ImageNet validation | top-1 accuracy: 78.9 -> 77.5 (-1.4) | Hurt |
| Add CutMix regularization | ReLabel-trained ResNet-50 · ImageNet validation | top-1 accuracy: 78.9 -> 80.2 | Helped |
| Add ImageNet-21K extra data on top of ReLabel and CutMix | ResNet-50 · ImageNet validation | top-1 accuracy: 80.2 -> 81.2 | Helped |
| Quantize label maps to 16-bit (ablation) | ReLabel on ResNet-50 · ImageNet validation | top-1 accuracy: 78.9 -> 78.8 | Hurt |
| Quantize label maps to 8-bit (ablation) | ReLabel on ResNet-50 · ImageNet validation | top-1 accuracy: 78.9 -> 78.6 | Hurt |
| Mix ReLabel annotations equally with original labels (ablation) | ReLabel-trained ResNet-50 · ImageNet validation | accuracy: 78.9 -> 78.3 | Hurt |
Closest papers by meaning
- Doomed to Re-Annotate, Forever: The ImageNet Story (NeurIPS 2026 · similarity 0.92)
- Multi-Label Learning From Single Positive Labels (CVPR 2021 · similarity 0.90)
- Is one annotation enough? - A data-centric image classification benchmark for noisy and ambiguous label estimation (NeurIPS 2022 · similarity 0.89)
- When does dough become a bagel? Analyzing the remaining mistakes on ImageNet (NeurIPS 2022 · similarity 0.89)
- CNN-RNN: A Unified Framework for Multi-Label Image Classification (CVPR 2016 · similarity 0.89)
- Multi-Label Image Recognition With Graph Convolutional Networks (CVPR 2019 · similarity 0.88)
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.