Research map / Label-efficient and robust learning / Recognition and long-tail
Between-Class Learning for Image Classification
Read the paperCheck 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 |
|---|---|---|---|
| Use BC image mixing with ratio targets instead of standard training | ResNeXt-101 (64×4)d · ImageNet-1K val, single-crop, 100 epochs | top-1 error (%): 20.4 -> 19.92 | Helped |
| Use BC image mixing with ratio targets instead of standard training | ResNeXt-101 (64×4)d · ImageNet-1K val, single-crop, 150 epochs | top-1 error (%): 20.44 -> 19.43 | Helped |
| Use BC image mixing with ratio targets instead of standard training | CIFAR-10 networks · CIFAR-10 | error rate (%): 6.07 ±0.04 -> 5.40 ±0.07 (11-layer CNN); 4.24 ± 0.06 -> 3.75 ±0.04 (ResNet-29); 3.54 ± 0.04 -> 2.79 ±0.06 (ResNeXt-29); | Helped |
| Use BC image mixing with ratio targets instead of standard training | CIFAR-100 networks · CIFAR-100 | error rate (%): 26.68 ±0.09 -> 24.28 ±0.11 (11-layer CNN); 20.18 ±0.07 -> 19.56 ±0.10 (ResNet-29); 16.99 ±0.06 -> 18.21 ±0.12 (ResNeXt-2 | Mixed |
| Use BC+ waveform-style mixing instead of simple BC mixing | 11-layer CNN, ResNet-29, and DenseNet-BC · CIFAR-10 | error rate (%): 5.40 ±0.07 -> 5.22 ±0.04 (11-layer CNN); 3.75 ±0.04 -> 3.55 ±0.03 (ResNet-29); 2.68 ±0.03 -> 2.57 ±0.06 (DenseNet-BC) | Helped |
| Use BC+ waveform-style mixing instead of simple BC mixing | ResNeXt-29 and Shake-Shake · CIFAR-10 | error rate (%): 2.79 ±0.06 -> 2.81 ±0.06 (ResNeXt-29); 2.38 ±0.04 -> 2.26 ±0.01 (Shake-Shake) | Mixed |
| Use BC+ waveform-style mixing instead of simple BC mixing | CIFAR-100 networks · CIFAR-100 | error rate (%): 24.28 ±0.11 -> 23.68 ±0.10 (11-layer CNN); 19.56 ±0.10 -> 19.41 ±0.07 (ResNet-29); 18.21 ±0.12 -> 17.93 ±0.09 (ResNeXt-2 | Mixed |
| Add per-image mean subtraction to simple BC mixing (ablation) | 11-layer CNN · CIFAR-10 | error rate (%): 5.40 -> 5.45 | No clear effect |
| Add variance normalization to per-image mean-subtracted mixing (ablation) | 11-layer CNN · CIFAR-10 | error rate (%): 5.45 -> 5.17 | Helped |
| Account for image energy differences in addition to mean subtraction and variance normalization (ablation) | 11-layer CNN · CIFAR-10 and CIFAR-100 | error rate (%): 5.17 -> 5.22 (CIFAR-10); 23.72 -> 23.68 (CIFAR-100) | No clear effect |
| Use ratio labels and KL loss instead of single-class or multilabel targets (ablation) | 11-layer CNN · CIFAR-10 and CIFAR-100 | error rate (%): 6.35 -> 5.22 (CIFAR-10); 27.28 -> 23.68 (CIFAR-100), versus Single; 6.05 -> 5.22 and 26.31 -> 23.68, versus Multi | Helped |
| Mix two images from different classes rather than sometimes mixing same-class images (ablation) | 11-layer CNN · CIFAR-10 and CIFAR-100 | error rate (%): 5.31 -> 5.22 (CIFAR-10); 23.79 -> 23.68 (CIFAR-100) | Helped |
Closest papers by meaning
- Learning from Between-class Examples for Deep Sound Recognition (ICLR 2018 · similarity 0.92)
- Provably Learning Diverse Features in Multi-View Data with Midpoint Mixup (ICML 2023 · similarity 0.86)
- CLEAR: Cumulative LEARning for One-Shot One-Class Image Recognition (CVPR 2018 · similarity 0.85)
- CleanNet: Transfer Learning for Scalable Image Classifier Training With Label Noise (CVPR 2018 · similarity 0.84)
- Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model (CVPR 2024 · similarity 0.84)
- Catch-Up Mix: Catch-Up Class for Struggling Filters in CNN (AAAI 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.