Home

Awesome

Awesome Long-Tailed Learning (TPAMI 2023)

Awesome PRs Welcome

We released Deep Long-Tailed Learning: A Survey and our codebase to the community. In this survey, we reviewed recent advances in long-tailed learning based on deep neural networks. Existing long-tailed learning studies can be grouped into three main categories (i.e., class re-balancing, information augmentation and module improvement), which can be further classified into nine sub-categories (as shown in the below figure). We also provided empirical analysis for several state-of-the-art methods by evaluating to what extent they address the issue of class imbalance. We concluded the survey by highlighting important applications of deep long-tailed learning and identifying several promising directions for future research.

After completing this survey, we decided to release our long-tailed learning resources and codebase, hoping to push the development of the community. If you have any questions or suggestions, please feel free to contact us.

<p align="center"> <img src="resources/Taxonomy2.png" width=1000> </p>

1. Type of Long-tailed Learning

SymbolSamplingCSLLATLAug
TypeRe-samplingClass-sensitive LearningLogit AdjustmentTransfer LearningData Augmentation
SymbolRLCDDTEnsembleother
TypeRepresentation LearningClassifier DesignDecoupled TrainingEnsemble LearningOther Types

2. Top-tier Conference Papers

2023

TitleVenueYearTypeCode
Long-tailed recognition by mutual information maximization between latent features and ground-truth labelsICML2023CSL,RLOfficial
Large language models struggle to learn long-tail knowledgeICML2023Aug
Feature directions matter: Long-tailed learning via rotated balanced representationICML2023RL
Wrapped Cauchy distributed angular softmax for long-tailed visual recognitionICML2023RL,CDOfficial
Rethinking image super resolution from long-tailed distribution learning perspectiveCVPR2023CSL
Transfer knowledge from head to tail: Uncertainty calibration under long-tailed distributionCVPR2023CSL,TLOfficial
Towards realistic long-tailed semi-supervised learning: Consistency is all you needCVPR2023CSL,TL,EnsembleOfficial
Global and local mixture consistency cumulative learning for long-tailed visual recognitionsCVPR2023CSL,RLOfficial
Long-tailed visual recognition via self-heterogeneous integration with knowledge excavationCVPR2023TL,EnsembleOfficial
Balancing logit variation for long-tailed semantic segmentationCVPR2023AugOfficial
Use your head: Improving long-tail video recognitionCVPR2023AugOfficial
FCC: Feature clusters compression for long-tailed visual recognitionCVPR2023RLOfficial
FEND: A future enhanced distribution-aware contrastive learning framework for long-tail trajectory predictionCVPR2023RL
SuperDisco: Super-class discovery improves visual recognition for the long-tailCVPR2023RL
Class-conditional sharpness-aware minimization for deep long-tailed recognitionCVPR2023DTOfficial
Balanced product of calibrated experts for long-tailed recognitionCVPR2023EnsembleOfficial
No one left behind: Improving the worst categories in long-tailed learningCVPR2023Ensemble
On the effectiveness of out-of-distribution data in self-supervised long-tail learningICLR2023Sampling,TL,AugOfficial
LPT: Long-tailed prompt tuning for image classificationICLR2023Sampling,TL,OtherOfficial
Long-tailed partial label learning via dynamic rebalancingICLR2023CSLOfficial
Delving into semantic scale imbalanceICLR2023CSL,RL
INPL: Pseudo-labeling the inliers first for imbalanced semi-supervised learningICLR2023TL
CUDA: Curriculum of data augmentation for long-tailed recognitionICLR2023AugOfficial
Long-tailed learning requires feature learningICLR2023RL
Decoupled training for long-tailed classification with stochastic representationsICLR2023RL,DT

2022

TitleVenueYearTypeCode
Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognitionNeurIPS2022CSL,EnsembleOfficial
SoLar: Sinkhorn label refinery for imbalanced partial-label learningNeurIPS2022CSLOfficial
Do we really need a learnable classifier at the end of deep neural network?NeurIPS2022RL,CD
Maximum class separation as inductive bias in one matrixNeurIPS2022CDOfficial
Escaping saddle points for effective generalization on class-imbalanced dataNeurIPS2022otherOfficial
Breadcrumbs: Adversarial class-balanced sampling for long-tailed recognitionECCV2022Sampling,Aug,DTOfficial
Constructing balance from imbalance for long-tailed image recognitionECCV2022Sampling,RLOfficial
Tackling long-tailed category distribution under domain shiftsECCV2022CSL,Aug,RLOfficial
Improving GANs for long-tailed data through group spectral regularizationECCV2022CSL,OtherOfficial
Learning class-wise visual-linguistic representation for long-tailed visual recognitionECCV2022TL,RLOfficial
Learning with free object segments for long-tailed instance segmentationECCV2022Aug
SAFA: Sample-adaptive feature augmentation for long-tailed image classificationECCV2022Aug,RL
On multi-domain long-tailed recognition, imbalanced domain generalization, and beyondECCV2022RLOfficial
Invariant feature learning for generalized long-tailed classificationECCV2022RLOfficial
Towards calibrated hyper-sphere representation via distribution overlap coefficient for long-tailed learningECCV2022RL,CDOfficial
Long-tailed instance segmentation using Gumbel optimized lossECCV2022CDOfficial
Long-tailed class incremental learningECCV2022DTOfficial
Identifying hard noise in long-tailed sample distributionECCV2022OtherOfficial
Relieving long-tailed instance segmentation via pairwise class balanceCVPR2022CSLOfficial
The majority can help the minority: Context-rich minority oversampling for long-tailed classificationCVPR2022TL,AugOfficial
Long-tail recognition via compositional knowledge transferCVPR2022TL,RL
BatchFormer: Learning to explore sample relationships for robust representation learningCVPR2022TL,RLOfficial
Nested collaborative learning for long-tailed visual recognitionCVPR2022RL,EnsembleOfficial
Long-tailed recognition via weight balancingCVPR2022DTOfficial
Class-balanced pixel-level self-labeling for domain adaptive semantic segmentationCVPR2022otherOfficial
Killing two birds with one stone: Efficient and robust training of face recognition CNNs by partial FCCVPR2022otherOfficial
Optimal transport for long-tailed recognition with learnable cost matrixICLR2022LA
Do deep networks transfer invariances across classes?ICLR2022TL,AugOfficial
Self-supervised learning is more robust to dataset imbalanceICLR2022RL

2021

TitleVenueYearTypeCode
Improving contrastive learning on imbalanced seed data via open-world samplingNeurIPS2021Sampling,TL, DCOfficial
Semi-supervised semantic segmentation via adaptive equalization learningNeurIPS2021Sampling,CSL,TL, AugOfficial
On model calibration for long-tailed object detection and instance segmentationNeurIPS2021LAOfficial
Label-imbalanced and group-sensitive classification under overparameterizationNeurIPS2021LA
Towards calibrated model for long-tailed visual recognition from prior perspectiveNeurIPS2021Aug, RLOfficial
Supercharging imbalanced data learning with energy-based contrastive representation transferNeurIPS2021Aug, TL, RLOfficial
VideoLT: Large-scale long-tailed video recognitionICCV2021SamplingOfficial
Exploring classification equilibrium in long-tailed object detectionICCV2021Sampling,CSLOfficial
GistNet: a geometric structure transfer network for long-tailed recognitionICCV2021Sampling,TL, DC
FASA: Feature augmentation and sampling adaptation for long-tailed instance segmentationICCV2021Sampling,CSL
ACE: Ally complementary experts for solving long-tailed recognition in one-shotICCV2021Sampling,EnsembleOfficial
Influence-Balanced Loss for Imbalanced Visual ClassificationICCV2021CSLOfficial
Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigationICCV2021TLOfficial
Self supervision to distillation for long-tailed visual recognitionICCV2021TLOfficial
Distilling virtual examples for long-tailed recognitionICCV2021TL
MosaicOS: A simple and effective use of object-centric images for long-tailed object detectionICCV2021TLOfficial
Parametric contrastive learningICCV2021RLOfficial
Distributional robustness loss for long-tail learningICCV2021RLOfficial
Learning of visual relations: The devil is in the tailsICCV2021DT
Image-Level or Object-Level? A Tale of Two Resampling Strategies for Long-Tailed DetectionICML2021SamplingOfficial
Self-Damaging Contrastive LearningICML2021TL,RLOfficial
Delving into deep imbalanced regressionICML2021OtherOfficial
Long-tailed multi-label visual recognition by collaborative training on uniform and re-balanced samplingsCVPR2021Sampling,Ensemble
Equalization loss v2: A new gradient balance approach for long-tailed object detectionCVPR2021CSLOfficial
Seesaw loss for long-tailed instance segmentationCVPR2021CSLOfficial
Adaptive class suppression loss for long-tail object detectionCVPR2021CSLOfficial
PML: Progressive margin loss for long-tailed age classificationCVPR2021CSL
Disentangling label distribution for long-tailed visual recognitionCVPR2021CSL,LAOfficial
Adversarial robustness under long-tailed distributionCVPR2021CSL,LA,CDOfficial
Distribution alignment: A unified framework for long-tail visual recognitionCVPR2021CSL,LA,DTOfficial
Improving calibration for long-tailed recognitionCVPR2021CSL,Aug,DTOfficial
CReST: A class-rebalancing self-training framework for imbalanced semi-supervised learningCVPR2021TLOfficial
Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual conceptsCVPR2021TLOfficial
RSG: A simple but effective module for learning imbalanced datasetsCVPR2021TL,AugOfficial
MetaSAug: Meta semantic augmentation for long-tailed visual recognitionCVPR2021AugOfficial
Contrastive learning based hybrid networks for long-tailed image classificationCVPR2021RL
Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervisionCVPR2021RL
Long-tail learning via logit adjustmentICLR2021LAOfficial
Long-tailed recognition by routing diverse distribution-aware expertsICLR2021TL,EnsembleOfficial
Exploring balanced feature spaces for representation learningICLR2021RL,DT

2020

TitleVenueYearTypeCode
Balanced meta-softmax for long-taield visual recognitionNeurIPS2020Sampling,CSLOfficial
Posterior recalibration for imbalanced datasetsNeurIPS2020LAOfficial
Long-tailed classification by keeping the good and removing the bad momentum causal effectNeurIPS2020LA,CDOfficial
Rethinking the value of labels for improving classimbalanced learningNeurIPS2020TL,RLOfficial
The devil is in classification: A simple framework for long-tail instance segmentationECCV2020Sampling,DT,EnsembleOfficial
Imbalanced continual learning with partitioning reservoir samplingECCV2020SamplingOfficial
Distribution-balanced loss for multi-label classification in long-tailed datasetsECCV2020CSLOfficial
Feature space augmentation for long-tailed dataECCV2020TL,Aug,DT
Learning from multiple experts: Self-paced knowledge distillation for long-tailed classificationECCV2020TL,EnsembleOfficial
Solving long-tailed recognition with deep realistic taxonomic classifierECCV2020CDOfficial
Learning to segment the tailCVPR2020Sampling,TLOfficial
BBN: Bilateral-branch network with cumulative learning for long-tailed visual recognitionCVPR2020Sampling,EnsembleOfficial
Overcoming classifier imbalance for long-tail object detection with balanced group softmaxCVPR2020Sampling,EnsembleOfficial
Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspectiveCVPR2020CSLOfficial
Equalization loss for long-tailed object recognitionCVPR2020CSLOfficial
Domain balancing: Face recognition on long-tailed domainsCVPR2020CSL
M2m: Imbalanced classification via majorto-minor translationCVPR2020TL,AugOfficial
Deep representation learning on long-tailed data: A learnable embedding augmentation perspectiveCVPR2020TL,Aug,RL
Inflated episodic memory with region self-attention for long-tailed visual recognitionCVPR2020RL
Decoupling representation and classifier for long-tailed recognitionICLR2020Sampling,CSL,RL,CD,DTOfficial

2019

TitleVenueYearTypeCode
Meta-weight-net: Learning an explicit mapping for sample weightingNeurIPS2019CSLOfficial
Learning imbalanced datasets with label-distribution-aware margin lossNeurIPS2019CSLOfficial
Dynamic curriculum learning for imbalanced data classificationICCV2019Sampling
Class-balanced loss based on effective number of samplesCVPR2019CSLOfficial
Striking the right balance with uncertaintyCVPR2019CSL
Feature transfer learning for face recognition with under-represented dataCVPR2019TL,Aug
Unequal-training for deep face recognition with long-tailed noisy dataCVPR2019RLOfficial
Large-scale long-tailed recognition in an open worldCVPR2019RLOfficial

2018

TitleVenueYearTypeCode
Large scale fine-grained categorization and domain-specific transfer learningCVPR2018TLOfficial

2017

TitleVenueYearTypeCode
Learning to model the tailNeurIPS2017CSL
Focal loss for dense object detectionICCV2017CSL
Range loss for deep face recognition with long-tailed training dataICCV2017RL
Class rectification hard mining for imbalanced deep learningICCV2017RL

2016

TitleVenueYearTypeCode
Learning deep representation for imbalanced classificationCVPR2016Sampling,RL
Factors in finetuning deep model for object detection with long-tail distributionCVPR2016CSL,RL

3. Benchmark Datasets

DatasetLong-tailed Task# Class# Training data# Test data
ImageNet-LTClassification1,000115,84650,000
CIFAR100-LTClassification10050,00010,000
Places-LTClassification36562,50036,500
iNaturalist 2018Classification8,142437,51324,426
LVIS v0.5Detection and Segmentation1,23057,00020,000
LVIS v1Detection and Segmentation1,203100,00019,800
VOC-LTMulti-label Classification201,1424,952
COCO-LTMulti-label Classification801,9095,000
VideoLTVideo Classification1,004179,35225,622

4. Our codebase

5. Empirical Studies

(1) Long-tailed benchmarking performance

<p align="center"> <img src="resources/Fig1.png" width=900> </p> <p align="center"> <img src="resources/Fig2.png" width=900> </p>

(2) More discussions on cost-sensitive losses

<p align="center"> <img src="resources/Fig3.png" width=500> </p>

5. Citation

If this repository is helpful to you, please cite our survey.

@article{zhang2023deep,
      title={Deep long-tailed learning: A survey},
      author={Zhang, Yifan and Kang, Bingyi and Hooi, Bryan and Yan, Shuicheng and Feng, Jiashi},
      journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
      year={2023},
      publisher={IEEE}
}

5. Other Resources