Decoupling Representation and Classifier for Long-Tailed Recognition

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.

PDF Abstract ICLR 2020 PDF ICLR 2020 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Long-tail learning with class descriptors AWA-LT LWS Per-Class Accuracy 73.4 # 4
Long-Tailed Accuracy 93.5 # 2
Long-tail Learning CIFAR-10-LT (ρ=10) LWS Error Rate 8.9 # 10
Long-tail Learning CIFAR-10-LT (ρ=10) cRT Error Rate 9.0 # 13
Long-tail learning with class descriptors CUB-LT LWS Per-Class Accuracy 53.1 # 3
Long-Tailed Accuracy 65.7 # 3
Long-tail Learning ImageNet-LT CB LWS Top-1 Accuracy 41.4 # 62
Long-tail learning with class descriptors ImageNet-LT-d LWS Per-Class Accuracy 49.9 # 3
Long-tail Learning iNaturalist 2018 CB-LWS Top-1 Accuracy 69.5% # 37
Long-tail Learning Places-LT CB LWS Top-1 Accuracy 37.6 # 25
Long-tail learning with class descriptors SUN-LT LWS Per-Class Accuracy 33.9 # 3
Long-Tailed Accuracy 40.2 # 2

Methods


No methods listed for this paper. Add relevant methods here