Research map / Label-efficient and robust learning / Recognition and long-tail
Long-Tailed Recognition via Information-Preservable Two-Stage Learning
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 two-stage long-tailed recognition method combining Balanced Negative Sampling (contrastive representation learning claimed equivalent to intra-class distance minimization) and IP-DPP (a determinantal point process sampler that keeps informative, low-confidence instances to rebalance classes).
Problem
Long-tailed image classification, where deep models are biased toward head classes and perform poorly on tail classes; existing one-stage and two-stage methods either lack good representations or lose information when rebalancing.
Approach
Stage 1 uses a contrastive NCE-style objective extended with same-class negatives/positives (BNS) to learn well-separated features; Stage 2 samples class-balanced subsets with a DPP whose kernel is built from per-sample correct-classification probabilities, so less-confidently classified (more informative) instances are preferentially selected.
Key results
- On CIFAR-10-LT the method reaches 76.4% overall accuracy, at least 2.6% above the best compared method.
- On ImageNet-LT it reaches 51.7% overall, 2.3% above DisA, and on iNaturalist 2018 it reaches 74.0%, 3.6% above SBCL.
- On CIFAR-100-LT (IF=100) it reaches 52.4% overall, 3.2% above DisA, though it trails OTmix by 10.7% on many-shot accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full BNS + IP-DPP two-stage method instead of the strongest baseline | Long-tailed recognition · CIFAR-10-LT | overall accuracy: 73.8 -> 76.4 | Helped |
| Use the full BNS + IP-DPP two-stage method instead of the strongest baseline | Long-tailed recognition · CIFAR-100-LT | overall accuracy: 49.2 -> 52.4 | Helped |
| Use the full BNS + IP-DPP two-stage method instead of the strongest baseline | Long-tailed recognition · ImageNet-LT | overall accuracy: 49.4 -> 51.7 | Helped |
| Use the full BNS + IP-DPP two-stage method instead of the strongest baseline | Long-tailed recognition · iNaturalist 2018 | overall accuracy: 70.4 -> 74.0 | Helped |
| Use BNS for representation learning instead of SBCL | Linear probing · CIFAR-10-LT | overall accuracy: 64.7 -> 68.2 | Helped |
| Use BNS for representation learning instead of SBCL | Linear probing · CIFAR-100-LT | overall accuracy: 44.2 -> 47.1 | Helped |
| Apply the full method instead of OTmix | Long-tailed recognition · CIFAR-10-LT | many-shot accuracy: 87.9 -> 82.0 | Hurt |
| Apply the full method instead of OTmix | Long-tailed recognition · CIFAR-100-LT | many-shot accuracy: 73.1 -> 62.4 | Hurt |
Surprising findings
- The method is clearly weaker than baselines on many-shot accuracy (e.g., 62.4% vs 73.1% for OTmix on CIFAR-100-LT), trading head-class performance for medium and tail gains.
Limitations
- Many-shot accuracy is lower than several baselines, indicating a head-class trade-off.
- The mutual-information-to-intra-class-distance equivalence is stated as proportionality in a theorem that rests on the authors' own derivation and assumptions such as independence across classes.
- The method uses a fixed small m for same-class sampling because minority classes have few samples.
Experimental setup
Datasets: CIFAR-10-LT CIFAR-100-LT ImageNet-LT iNaturalist 2018
Benchmarks: CIFAR-10-LT CIFAR-100-LT ImageNet-LT iNaturalist 2018
Metrics: many-shot accuracy medium-shot accuracy few-shot accuracy overall accuracy linear probing accuracy
Baselines: Focal Loss LDAM Loss tau-norm RIDE KCL TSC SBCL OTmix DisA
Builds on: SBCL Noise Contrastive Estimation determinantal point process L-ensemble
Closest papers by meaning
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.89)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.89)
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.89)
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.88)
- Exploring Weight Balancing on Long-Tailed Recognition Problem (ICLR 2024 · similarity 0.87)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · 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.