Research map / Label-efficient and robust learning / Recognition and long-tail
Decoupled Contrastive Learning for Long-Tailed Recognition
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 Decoupled Contrastive Learning (DSCL + patch-based self distillation PBSD) to fix biased positive-pair optimization in supervised contrastive learning for long-tailed image classification.
Problem
Supervised contrastive learning on long-tailed data treats augmented views and same-class images equally, producing imbalanced intra-category gradients and under-representation of tail classes.
Approach
Decouples the two types of positives in SCL with separate weights (alpha) so their gradient ratio and optimal probability are independent of class cardinality, and adds patch-based self distillation that transfers patch-to-image similarity relations from head to tail classes.
Key results
- On ImageNet-LT, DSCL+PBSD reaches 57.7% top-1 with ResNet-50, improving SCL baseline (51.2%) by 6.5 points.
- Combined with RIDE, the method reaches 59.7% on ImageNet-LT and 74.9% overall on iNaturalist 2018, matching NCL's 74.9%.
- Achieves 42.4% overall on Places-LT with a single model, the best reported in its comparison table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace SCL with DSCL (ablation) | SCL on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 51.2 -> 52.6 | Helped |
| Add PBSD to DSCL (ablation) | DSCL on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 52.6 -> 57.7 | Helped |
| Use global-view features instead of patch-based features in PBSD (ablation) | DSCL + PBSD on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 57.7 -> 56.2 | Hurt |
| Replace PBSD with the multi-crop trick (ablation) | DSCL + PBSD on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 57.7% to 56.1% | Hurt |
| Set α to 0, removing the augmented-view positive (ablation) | DSCL + PBSD on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 57.7% to 56.8% | Hurt |
| Set α to 1, leaving only the augmented-view positive (ablation) | DSCL + PBSD on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 39.8% | Hurt |
| Increase patch-based features per anchor from 1 to 5 (ablation) | DSCL + PBSD on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 55.0% to 57.7% | Helped |
| Set second-stage classifier learning rate to 2.5 (ablation) | Our on ImageNet-LT · ImageNet-LT | top-1 accuracy (Few): 35.4% to 38.7% | Helped |
| Set second-stage classifier learning rate to 2.5 (ablation) | Our on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 57.7 -> 57.4 | Hurt |
| Combine Our with RIDE | Our on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 57.7 -> 59.7 | Helped |
| Combine Our with RIDE | Our on iNaturalist 2018 · iNaturalist 2018 | top-1 accuracy (Overall): 72.0 -> 74.9 | Helped |
| Use the fair-comparison training setup for PaCo (same augmentation and epochs) | PaCo on ImageNet-LT · ImageNet-LT | top-1 accuracy (Overall): 57.0% to 53.6% | Hurt |
Surprising findings
- Setting alpha=1 (pure self-supervised-style pulling of augmentations) drops accuracy to 39.8%, while alpha=0 still beats the SCL baseline.
- Replacing patch-based features with global-view features in the distillation loss costs about 1.5 points overall accuracy.
Limitations
- Gains on Many-shot split are sometimes accompanied by reduced few-shot accuracy depending on second-stage learning rate, so per-split results are sensitive to hyperparameters.
- Best ImageNet-LT result relies on ensembling with RIDE for the 59.7% figure; the single-model result is 57.7%.
- Fairness comparisons required retraining PaCo under the same augmentation and epochs, showing reported baselines use differing setups.
Experimental setup
Datasets: ImageNet-LT iNaturalist 2018 Places-LT
Benchmarks: ImageNet-LT iNaturalist 2018 Places-LT
Metrics: top-1 accuracy many-shot accuracy medium-shot accuracy few-shot accuracy
Baselines: CE Balanced Softmax cRT DisAlign BatchFormer KCL PaCo TSC BCL RIDE NCL SADE
Builds on: Supervised Contrastive Loss MoCoV2 KCL TSC RIDE
Closest papers by meaning
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.91)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 2024 · similarity 0.89)
- Decoupled Optimisation for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
- DiffuLT: Diffusion for Long-tail Recognition Without External Knowledge (NeurIPS 2024 · similarity 0.87)
- Long-Tailed Recognition via Information-Preservable Two-Stage Learning (NeurIPS 2025 · 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.