Research map / Label-efficient and robust learning / Recognition and long-tail
Decoupled Optimisation for Long-Tailed Visual 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 Optimisation (DO), a multi-stage framework that splits ResNet parameters into groups each specialised for a many/medium/few class group, combined with instance-level test-time learning to aggregate sub-models for long-tailed visual recognition.
Problem
Conventional training on long-tailed datasets biases model parameters toward head (many-shot) classes, leaving medium and few classes under-served.
Approach
Decouples network parameters via Taylor-based importance into subsets, each trained with group-preferred sampling and a compensation loss for one class group across stages; inference aggregates sub-models with instance-level test-time self-supervised weights.
Key results
- DO achieves 60.4% on ImageNet-LT, 75.8% on iNaturalist2018, and 42.8% on Places-LT, outperforming the listed multi-expert methods on these datasets.
- On ImageNet-LT, DO reaches 52.2% few-group accuracy in the three-stage setting versus 36.0% for RIDE, and ablation shows instance-level test-time learning adds about 2-3 points over group-level aggregation.
- DO is about 3.4% better than PaCo on ImageNet-LT, and on CIFAR100-LT it reaches 58.2% (β=50) and 53.8% (β=100).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DO instead of BalPoE | ImageNet-LT classification · ImageNet-LT test | accuracy (All): 59.7 -> 60.4 | Helped |
| Use DO instead of SHIKE | iNaturalist2018 classification · iNaturalist2018 test | accuracy (All): 75.4 -> 75.8 | Helped |
| Use DO instead of SHIKE | Places-LT classification · Places-LT test | accuracy (All): 41.9 -> 42.8 | Helped |
| Use DO instead of BalPoE | CIFAR100-LT β=50 classification · CIFAR100-LT β=50 test | accuracy (All): 58.7 -> 58.2 | Hurt |
| Use DO instead of BalPoE | CIFAR100-LT β=100 classification · CIFAR100-LT β=100 test | accuracy (All): 54.7 -> 53.8 | Hurt |
| Use three stages instead of two (ablation) | DO on ImageNet-LT · ImageNet-LT test | accuracy (Avg): 50.3 -> 60.4 | Helped |
| Split the medium group into two groups and use four stages (ablation) | DO on ImageNet-LT · ImageNet-LT test | accuracy (Avg): 60.4 -> 60.8 | Helped |
| Order stages as many, few, then medium instead of many, medium, then few (ablation) | DO on ImageNet-LT · ImageNet-LT test | accuracy by class group: 67.0 / 57.3 / 52.2 -> 67.1 / 53.6 / 51.0 | Mixed |
| Replace instance-level test-time learning with average aggregation (ablation) | DO on ImageNet-LT · ImageNet-LT test | accuracy: 57.9 -> 60.4 | Hurt |
| Increase the medium-enhanced sampling ratio (ablation) | DO on ImageNet-LT and CIFAR100-LT · ImageNet-LT and CIFAR100-LT | accuracy by class group | Mixed |
Surprising findings
- Learning the many group first and then few group (default) outperformed a medium-first order, which contradicted the intuition that medium classes benefit most from early focus; the paper attributes this to later stages having more free parameters.
Limitations
- Instance-level test-time learning increases inference cost compared with simpler aggregation.
- Hyperparameters such as the medium-enhanced ratio and pruning ratio are tuned on the validation set, which is also used for benchmarking on datasets without a separate test split.
Experimental setup
Datasets: ImageNet-LT iNaturalist2018 Places-LT CIFAR100-LT
Benchmarks: ImageNet-LT iNaturalist2018 Places-LT CIFAR100-LT
Metrics: top-1 accuracy many/medium/few group accuracy
Baselines: CE LWS BSCE MiSLAS LADE CMO RSG PaCo TSC GCL CC-SAM RIDE
Builds on: Decoupling representation and classifier (Kang et al. 2019) SADE gradual pruning (Zhu and Gupta 2017) PaCo RIDE
Closest papers by meaning
- Decoupling Representation and Classifier for Long-Tailed Recognition (ICLR 2020 · similarity 0.91)
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.88)
- Improving Visual Prompt Tuning by Gaussian Neighborhood Minimization for Long-Tailed Visual Recognition (NeurIPS 2024 · similarity 0.88)
- Decoupled Contrastive Learning for Long-Tailed Recognition (AAAI 2024 · similarity 0.87)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
- OFBD: Object-Focused Background Debiasing for Long-Tailed Learning (NeurIPS 2026 · 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.