Research map / Label-efficient and robust learning / Recognition and long-tail

OFBD: Object-Focused Background Debiasing for Long-Tailed Learning

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Decoupled and calibrated long-tail training · 0 citations (OpenAlex)

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 OFBD, a plug-in long-tailed learning framework that combines RL-selected foreground-guided CutMix with a parameter-free background-guided feature rectification to reduce background bias and improve tail-class accuracy.

Problem

Long-tailed visual recognition degrades on tail classes; the paper argues this is partly caused by background-biased representations and optimization, not only by sample scarcity.

Approach

Break foreground-background co-occurrence by an RL-learned selector that chooses foreground regions for CutMix (preserving target content while replacing backgrounds), and rectify features by down-weighting background-biased spatial locations using a parameter-free SimAM-inspired score.

Key results

• On CIFAR-100-LT (r=100), adding OFBD to CE raises overall accuracy from 38.32% to 44.83% (+6.51), with Few-class accuracy rising from 9.14% to 23.93%.
• Adding OFBD to BCL on CIFAR-100-LT r=100 improves accuracy from 51.84% to 54.19% (+2.35); the RL selector beats random, SAM-based and CAM-based CutMix region selection by +2.02, +1.76 and +1.56 points.
• On ImageNet-LT with ResNet-50, OFBD improves CE from 44.80% to 47.39%; on iNaturalist 2018 CE improves from 65.95% to 69.76%.

Surprising findings

• Accurate foreground localization via pretrained CAM or SAM is not necessarily the best CutMix mixing region; a learned RL selector performs better.
• OFBD slightly reduces Many-class accuracy on some strong baselines (e.g., -0.06 on BCL with r=100 CIFAR-100-LT Many group, and -1.23 to -3.73 on Many classes at higher imbalance ratios).

Limitations

• Performance on highly complex scene domains such as medical imaging is unexplored.
• Several theoretical claims rest on assumed inequalities and heuristic derivations rather than rigorous proofs.
• Gains on strong baselines are often small (under 1 point on some ImageNet-LT and iNaturalist settings).

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: top-1 accuracy many/medium/few-shot accuracy background gradient ratio shift BG AUROC FG-only/BG-only/BG-swap accuracy

Baselines: CE LDAM-DRW BBN BCL SBCL GBG MKP GCL ViT DeiT-LT SaliencyMix SnapMix

Builds on: CutMix GRPO SimAM LaSt-ViT BCL

Closest papers by meaning

• SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection (CVPR 2025 · similarity 0.87)
• Adaptive Hierarchical Representation Learning for Long-Tailed Object Detection (CVPR 2022 · similarity 0.87)
• Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
• Bag of Tricks for Long-Tailed Visual Recognition with Deep Convolutional Neural Networks (AAAI 2021 · similarity 0.87)
• Decoupled Optimisation for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
• Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 2024 · similarity 0.87)