TasksSotADatasetsPapersMethodsSubmitAbout
Papers With Code 2

A community resource for machine learning research: papers, code, benchmarks, and state-of-the-art results.

Explore

Notable BenchmarksAll SotADatasetsPapersMethods

Community

Submit ResultsAbout

Data sourced from the PWC Archive (CC-BY-SA 4.0). Built by the community, for the community.

Papers/Decoupling Representation and Classifier for Long-Tailed R...

Decoupling Representation and Classifier for Long-Tailed Recognition

Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, Yannis Kalantidis

2019-10-21ICLR 2020 1Representation LearningLong-tail LearningLong-tail learning with class descriptorsTransfer LearningGeneral ClassificationClassification
PaperPDFCode(official)CodeCodeCode

Abstract

The long-tail distribution of the visual world poses great challenges for deep learning based classification models on how to handle the class imbalance problem. Existing solutions usually involve class-balancing strategies, e.g., by loss re-weighting, data re-sampling, or transfer learning from head- to tail-classes, but most of them adhere to the scheme of jointly learning representations and classifiers. In this work, we decouple the learning procedure into representation learning and classification, and systematically explore how different balancing strategies affect them for long-tailed recognition. The findings are surprising: (1) data imbalance might not be an issue in learning high-quality representations; (2) with representations learned with the simplest instance-balanced (natural) sampling, it is also possible to achieve strong long-tailed recognition ability by adjusting only the classifier. We conduct extensive experiments and set new state-of-the-art performance on common long-tailed benchmarks like ImageNet-LT, Places-LT and iNaturalist, showing that it is possible to outperform carefully designed losses, sampling strategies, even complex modules with memory, by using a straightforward approach that decouples representation and classification. Our code is available at https://github.com/facebookresearch/classifier-balancing.

Results

TaskDatasetMetricValueModel
Image ClassificationPlaces-LTTop-1 Accuracy37.6CB LWS
Image ClassificationCIFAR-10-LT (ρ=10)Error Rate8.9LWS
Image ClassificationCIFAR-10-LT (ρ=10)Error Rate9cRT
Image ClassificationImageNet-LTTop-1 Accuracy41.4CB LWS
Image ClassificationCUB-LTLong-Tailed Accuracy65.7LWS
Image ClassificationCUB-LTPer-Class Accuracy53.1LWS
Image ClassificationAWA-LTLong-Tailed Accuracy93.5LWS
Image ClassificationAWA-LTPer-Class Accuracy73.4LWS
Image ClassificationSUN-LTLong-Tailed Accuracy40.2LWS
Image ClassificationSUN-LTPer-Class Accuracy33.9LWS
Image ClassificationImageNet-LT-dPer-Class Accuracy49.9LWS
Few-Shot Image ClassificationPlaces-LTTop-1 Accuracy37.6CB LWS
Few-Shot Image ClassificationCIFAR-10-LT (ρ=10)Error Rate8.9LWS
Few-Shot Image ClassificationCIFAR-10-LT (ρ=10)Error Rate9cRT
Few-Shot Image ClassificationImageNet-LTTop-1 Accuracy41.4CB LWS
Few-Shot Image ClassificationCUB-LTLong-Tailed Accuracy65.7LWS
Few-Shot Image ClassificationCUB-LTPer-Class Accuracy53.1LWS
Few-Shot Image ClassificationAWA-LTLong-Tailed Accuracy93.5LWS
Few-Shot Image ClassificationAWA-LTPer-Class Accuracy73.4LWS
Few-Shot Image ClassificationSUN-LTLong-Tailed Accuracy40.2LWS
Few-Shot Image ClassificationSUN-LTPer-Class Accuracy33.9LWS
Few-Shot Image ClassificationImageNet-LT-dPer-Class Accuracy49.9LWS
Generalized Few-Shot ClassificationPlaces-LTTop-1 Accuracy37.6CB LWS
Generalized Few-Shot ClassificationCIFAR-10-LT (ρ=10)Error Rate8.9LWS
Generalized Few-Shot ClassificationCIFAR-10-LT (ρ=10)Error Rate9cRT
Generalized Few-Shot ClassificationImageNet-LTTop-1 Accuracy41.4CB LWS
Generalized Few-Shot ClassificationCUB-LTLong-Tailed Accuracy65.7LWS
Generalized Few-Shot ClassificationCUB-LTPer-Class Accuracy53.1LWS
Generalized Few-Shot ClassificationAWA-LTLong-Tailed Accuracy93.5LWS
Generalized Few-Shot ClassificationAWA-LTPer-Class Accuracy73.4LWS
Generalized Few-Shot ClassificationSUN-LTLong-Tailed Accuracy40.2LWS
Generalized Few-Shot ClassificationSUN-LTPer-Class Accuracy33.9LWS
Generalized Few-Shot ClassificationImageNet-LT-dPer-Class Accuracy49.9LWS
Long-tail LearningPlaces-LTTop-1 Accuracy37.6CB LWS
Long-tail LearningCIFAR-10-LT (ρ=10)Error Rate8.9LWS
Long-tail LearningCIFAR-10-LT (ρ=10)Error Rate9cRT
Long-tail LearningImageNet-LTTop-1 Accuracy41.4CB LWS
Long-tail LearningCUB-LTLong-Tailed Accuracy65.7LWS
Long-tail LearningCUB-LTPer-Class Accuracy53.1LWS
Long-tail LearningAWA-LTLong-Tailed Accuracy93.5LWS
Long-tail LearningAWA-LTPer-Class Accuracy73.4LWS
Long-tail LearningSUN-LTLong-Tailed Accuracy40.2LWS
Long-tail LearningSUN-LTPer-Class Accuracy33.9LWS
Long-tail LearningImageNet-LT-dPer-Class Accuracy49.9LWS
Generalized Few-Shot LearningPlaces-LTTop-1 Accuracy37.6CB LWS
Generalized Few-Shot LearningCIFAR-10-LT (ρ=10)Error Rate8.9LWS
Generalized Few-Shot LearningCIFAR-10-LT (ρ=10)Error Rate9cRT
Generalized Few-Shot LearningImageNet-LTTop-1 Accuracy41.4CB LWS
Generalized Few-Shot LearningCUB-LTLong-Tailed Accuracy65.7LWS
Generalized Few-Shot LearningCUB-LTPer-Class Accuracy53.1LWS
Generalized Few-Shot LearningAWA-LTLong-Tailed Accuracy93.5LWS
Generalized Few-Shot LearningAWA-LTPer-Class Accuracy73.4LWS
Generalized Few-Shot LearningSUN-LTLong-Tailed Accuracy40.2LWS
Generalized Few-Shot LearningSUN-LTPer-Class Accuracy33.9LWS
Generalized Few-Shot LearningImageNet-LT-dPer-Class Accuracy49.9LWS

Related Papers

Touch in the Wild: Learning Fine-Grained Manipulation with a Portable Visuo-Tactile Gripper2025-07-20RaMen: Multi-Strategy Multi-Modal Learning for Bundle Construction2025-07-18Spectral Bellman Method: Unifying Representation and Exploration in RL2025-07-17Boosting Team Modeling through Tempo-Relational Representation Learning2025-07-17Disentangling coincident cell events using deep transfer learning and compressive sensing2025-07-17Adversarial attacks to image classification systems using evolutionary algorithms2025-07-17Similarity-Guided Diffusion for Contrastive Sequential Recommendation2025-07-16Are encoders able to learn landmarkers for warm-starting of Hyperparameter Optimization?2025-07-16