cs.CVSep 29, 2026

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

Authors: Shenghan Chen, Yiming Liu, Zhipeng Deng, Haolin Wang, Jiale Zhou, Zhijian Wu, Xiankai Lu, Yafei Ou, +1 more

Organizations: Westlake University, Hangzhou, China · Shandong University, Jinan, China · Hokkaido University, Sapporo, Japan · RIKEN, Japan

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Balancing performance trade-offs on long-tailed data distributions remains a long-standing challenge in visual recognition. Existing methods mainly improve tail classes through re-balancing, representation learning, or data augmentation, but the underlying cause of tail class degradation is still insufficiently explored. In this paper, we find that standard long-tailed training induces background-biased representation and optimization: tail classes suffer larger background distribution shifts and become increasingly driven by background gradients. This reveals that tail degradation is not merely caused by insufficient samples, but also by the learning of irrelevant background features. To tackle this issue, we propose Object-Focused Background Debiasing (OFBD), a framework that mitigates background bias from both distribution and optimization perspectives. Specifically, Foreground-guided CutMix preserves target-related foregrounds while diversifying complementary backgrounds, and Background-guided Feature Rectification suppresses background-biased features without learnable parameters or additional training. Extensive experiments show that our method improves overall accuracy, achieves significant tail-class gains, and can serve as a plug-in for mainstream long-tailed methods without external data or pretrained recognition models. The code is available at: https://ofbd-neurips2026-longtail-learning.github.io/

Figures & tables

Figure 1 : Motivation of OFBD. Tail-class degradation stems from both sample scarcity and background bias in distribution and optimization. (a) Head classes possess diverse, class-agnostic backgrounds (BGs), making background representations (BG-Rep.) uninformative for classification. Conversely, tail classes often co-occur with class-correlated BGs ( e.g. , snow fox in snowy scenes), which makes BG-Rep. a predictive shortcut and encourages reliance on BG cues over sparse foregrounds. Gray dashed lines denote BG-based decision boundaries. (b) Optimization-level background bias. “Head” and “Tail” represent the evaluation results under the long-tailed setting, while “All” indicates the overall result on the full balanced dataset. (c) Attention maps reveal that tail predictions co-activate foregrounds and backgrounds, whereas head predictions remain object-focused.
Figure 2 : Overview of OFBD. The pink box denotes OFBD as a plug-in module for mainstream long-tailed models. (a) FG-CutMix selects target-related foregrounds from random regions and replaces complementary backgrounds, while BFR estimates background-aware scores to rectify features before classification. (b) The region selector πθ\pi_{\theta} is optimized via reinforcement learning, and a frozen reference selector πref\pi_{\mathrm{ref}} stabilizes the training of the selector.
 CIFAR-100-LTCIFAR-10-LTStatistic (r=100)(r=100)
Methodr=100 ↑\uparrowr=50 ↑\uparrowr=10 ↑\uparrowr=100 ↑\uparrowr=50 ↑\uparrowr=10 ↑\uparrowMany ↑\uparrowMed. ↑\uparrowFew ↑\uparrow
CE38.3243.9455.7170.4574.8286.4465.1937.069.14
+OFBD44.83_{{\color[rgb]{0,0.6,0}\uparrow\,6.51}}47.79_{{\color[rgb]{0,0.6,0}\uparrow\,3.85}}58.22_{{\color[rgb]{0,0.6,0}\uparrow\,2.51}}74.33_{{\color[rgb]{0,0.6,0}\uparrow\,3.88}}78.68_{{\color[rgb]{0,0.6,0}\uparrow\,3.86}}87.30_{{\color[rgb]{0,0.6,0}\uparrow\,0.86}}65.48_{{\color[rgb]{0,0.6,0}\uparrow\,0.29}}42.09_{{\color[rgb]{0,0.6,0}\uparrow\,5.03}}23.93_{{\color[rgb]{0,0.6,0}\uparrow\,14.79}}
LDAM-DRW [ 4 ] (NeurIPS’19)42.0946.5758.6477.0681.0988.0261.5141.6720.22
+OFBD45.81_{{\color[rgb]{0,0.6,0}\uparrow\,3.72}}49.55_{{\color[rgb]{0,0.6,0}\uparrow\,2.98}}59.09_{{\color[rgb]{0,0.6,0}\uparrow\,0.45}}79.61_{{\color[rgb]{0,0.6,0}\uparrow\,2.55}}82.43_{{\color[rgb]{0,0.6,0}\uparrow\,1.34}}88.76_{{\color[rgb]{0,0.6,0}\uparrow\,0.74}}64.93_{{\color[rgb]{0,0.6,0}\uparrow\,3.42}}45.81_{{\color[rgb]{0,0.6,0}\uparrow\,4.14}}26.61_{{\color[rgb]{0,0.6,0}\uparrow\,6.39}}
BBN [ 107 ] (CVPR’20)42.4547.1359.2379.7881.2488.3242.6350.7032.60
Table 1 : Comparisons on CIFAR-LT [ 4 ] datasets with varying imbalance ratios ( rr ). Results are reported across all classes as well as different groups based on training sample sizes ("Many", "Med", and "Few"). For each method, the first row reports the original baseline accuracy, and the second row shows the performance after applying OFBD.
Table 4
Figure 3 : Visualization of background debiasing on CIFAR100-LT ( r=100r=100 ). (a) Pretrained-CAM attention maps of CE and CE+OFBD on three representative head, medium, and tail classes, showing that OFBD reduces background co-activation. More visualization results are provided in Appendix C.1 . (b) Background gradient ratio (Eq. 2 ) shift during training. OFBD reduces optimization-level background bias (Sec. 3.3 ), especially for tail classes.
Table 6
Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4 : Extended pretrained-CAM attention maps of representative Head, Medium, and Tail classes under CE, CE+OFBD, BCL, and BCL+OFBD. OFBD reduces background co-activation and makes attention more object-focused across different baselines and class-frequency groups.
Table 8
Table 9
Figure 5 : Empirical verification of background bias on CIFAR-100-LT with imbalance ratio 100. Tail classes exhibit both larger background distribution shift and stronger background-gradient-ratio shift than head classes, validating the distribution-level and optimization-level bias analyzed in Sec. 3 .
Figure 6 : Efficiency analysis of OFBD on CIFAR-100-LT with imbalance ratio 100100 . (a) Accuracy and per-epoch (s) computation cost as a function of the candidate pool size KK . Accuracy first improves and then slightly decreases for large KK , while per-epoch cost rises monotonically. K=6K=6 is selected as the default for a good efficiency-accuracy trade-off. (b) Efficiency-accuracy comparison with representative long-tailed methods, where per-epoch computation cost is measured in seconds. “Ours” denotes BCL+OFBD.
Figure 12
Figure 9 : Further analysis of FG-CutMix on CIFAR-100-LT with imbalance ratio 100100 . (a) FG-CutMix reduces tail-class background-gradient-ratio shift compared with standard CutMix, indicating weaker background-driven optimization. (b) The RL loss of the foreground selector remains bounded and gradually stabilizes during training, indicating stable optimization of the RL-based selector.
Figure 10 : Eigen Spectral Density of Hessian for head and tail classes of ResNet models trained with CE, BCL, and BCL+OFBD on CIFAR-100 LT respectively. A smaller λmax⁡\lambda_{\max} and Tr(H)Tr(H) generally indicate a flatter loss landscape.
Table 15
Figure 11 : Visualization of loss landscape for head and tail classes of ResNet models trained with CE, BCL, and BCL+OFBD on CIFAR-100 LT respectively.

Explore similar work

CardsList
May 24, 2026cs.LG

OSDTW: Optimal Shared Depth and Task Weighting for Long-Tailed Recognition

Long-tailed recognition suffers from a persistent head--tail trade-off: improving tail performance often degrades head accuracy and can increase training instability. Despite strong empirical results from re-weighting, decoupled training, and multi-expert methods, key design choices about representation sharing between head and tail classes and supervision weighting across class groups remain largely heuristic. In this work, we propose OSDTW, a principled task-decomposition framework that partitions the original single-label recognition problem into a head task and a tail task, implemented with a shared encoder and task-specific decoders. To handle the mutual exclusivity and statistical dependence between the two label groups, we introduce a factorized model and show that the resulting Kullback--Leibler divergence-based generalization error can be written as the sum of task-wise terms up to an additive constant, yielding a well-defined task-wise objective. We further develop a three-stage training pipeline: independent task training to estimate task-wise optima and the Fisher information matrix, weighted joint training to learn a shared encoder, and branch assembly to construct the final decoupled model. Under a block-diagonal Fisher approximation, we derive a computable second-order expansion of the expected generalization error, decomposing it into encoder variance, encoder bias, and decoder variance. This bias--variance decomposition provides a computable proxy to select the shared depth and task weights, enabling efficient hyper-parameter search. Experiments on standard long-tailed benchmarks demonstrate the effectiveness of the proposed approach over strong baselines.

Chang Chu, Qingyue Zhang, Shao-Lun Huang +1

Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Shenzhen Zkosemi Semiconductor Technology Co., Ltd, Shenzhen, China

Multi--Task LearningBias-Variance Decomposition
AskPDF
May 2, 2026cs.CV

Decision Boundary-aware Generation for Long-tailed Learning

Long-tailed data bias decision boundaries toward head classes and degrade tail class accuracy. Diffusion-based generative augmentation address this problem by generating additional data, while head-to-tail transfer further mitigate the generator bias inherit from long-tailed dataset. However, we show that while head-to-tail transfer helps balance the decision space of the classifier, it also induces latent non-local feature mixing that entangles inter-class features, causing decision boundary overlap and tail class distribution shift. To address this, we first identify the problem of boundary ambiguity and then propose Decision Boundary-aware Generation (DBG) framework, which promotes near-boundary representation learning by generating informative near-boundary samples. Overall, DBG rebalances the long-tailed dataset while yielding more separable decision space for long-tailed learning. Across standard long-tailed benchmarks, DBG consistently improves tail class and overall accuracy with less inter-class overlap. The code of DBG is available at https://github.com/keepdigitalabc-svg/DBG.

Jiacheng Yang, Ruichi Zhang, Chikai Shang +5

Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, Xiamen, China · College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China · Department of Statistical Science, University College London, London, UK +1

Imbalanced ClassificationGenerative ModelsDistribution Shifts
AskPDF
May 2, 2026cs.CV

CUE: Concept-Aware Multi-Label Expansion to Mitigate Concept Confusion in Long-Tailed Learning

Long-tailed distributions are common in real-world recognition tasks, where a few head classes have many samples while most tail classes have very few. Recently, fine-tuning foundation models for long-tailed learning has gained attention due to their excellent performance. However, most existing methods focus solely on mitigating long-tailed distribution bias while overlooking concept confusion caused by the long-tailed distribution. In this paper, we study this problem and attribute it to the mutual exclusivity of single-label supervision under long-tailed distributions, which suppresses feature sharing among related classes and amplifies the dominance of head classes, leading to disrupted inter-class discriminability. To address this, we propose CUE, Concept-aware mUlti-label Expansion, which introduces multi-label concept signals to preserve disrupted inter-class relationships. Specifically, CUE constructs concept sets by (i) extracting instance-level visual cues from zero-shot CLIP and (ii) generating class-level semantic cues with LLM; the two cues are incorporated via separately weighted Binary Logit-Adjustment (BLA) auxiliary losses and jointly optimized with the baseline Logit-Adjustment (LA) loss. Experiments on several long-tailed benchmarks, CUE achieves balanced and strong performance, surpassing recent state-of-the-art methods. Code is available at: https://github.com/zhangruichi/CUE.

Ruichi Zhang, Chikai Shang, Jiacheng Yang +4

Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China · College of Computer Science and Software Engineering, Shenzhen University, China · Institute of High Performance Computing, A*STAR, Singapore

Multi-Label ClassificationLong-Tailed Distribution
AskPDF