Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, Stella X. Yu
Natural data are often long-tail distributed over semantic classes. Existing recognition methods tackle this imbalanced classification by placing more emphasis on the tail data, through class re-balancing/re-weighting or ensembling over different data groups, resulting in increased tail accuracies but reduced head accuracies. We take a dynamic view of the training data and provide a principled model bias and variance analysis as the training data fluctuates: Existing long-tail classifiers invariably increase the model variance and the head-tail model bias gap remains large, due to more and larger confusion with hard negatives for the tail. We propose a new long-tailed classifier called RoutIng Diverse Experts (RIDE). It reduces the model variance with multiple experts, reduces the model bias with a distribution-aware diversity loss, reduces the computational cost with a dynamic expert routing module. RIDE outperforms the state-of-the-art by 5% to 7% on CIFAR100-LT, ImageNet-LT and iNaturalist 2018 benchmarks. It is also a universal framework that is applicable to various backbone networks, long-tailed algorithms, and training mechanisms for consistent performance gains. Our code is available at: https://github.com/frank-xwang/RIDE-LongTailRecognition.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Image Classification | ImageNet-LT | Top-1 Accuracy | 56.4 | RIDE (ResNeXt-50) |
| Image Classification | ImageNet-LT | Top-1 Accuracy | 54.9 | RIDE (ResNet-50) |
| Image Classification | CIFAR-100-LT (ρ=100) | Error Rate | 50.9 | RIDE+distill |
| Image Classification | CIFAR-100-LT (ρ=100) | Error Rate | 52 | RIDE |
| Few-Shot Image Classification | ImageNet-LT | Top-1 Accuracy | 56.4 | RIDE (ResNeXt-50) |
| Few-Shot Image Classification | ImageNet-LT | Top-1 Accuracy | 54.9 | RIDE (ResNet-50) |
| Few-Shot Image Classification | CIFAR-100-LT (ρ=100) | Error Rate | 50.9 | RIDE+distill |
| Few-Shot Image Classification | CIFAR-100-LT (ρ=100) | Error Rate | 52 | RIDE |
| Generalized Few-Shot Classification | ImageNet-LT | Top-1 Accuracy | 56.4 | RIDE (ResNeXt-50) |
| Generalized Few-Shot Classification | ImageNet-LT | Top-1 Accuracy | 54.9 | RIDE (ResNet-50) |
| Generalized Few-Shot Classification | CIFAR-100-LT (ρ=100) | Error Rate | 50.9 | RIDE+distill |
| Generalized Few-Shot Classification | CIFAR-100-LT (ρ=100) | Error Rate | 52 | RIDE |
| Long-tail Learning | ImageNet-LT | Top-1 Accuracy | 56.4 | RIDE (ResNeXt-50) |
| Long-tail Learning | ImageNet-LT | Top-1 Accuracy | 54.9 | RIDE (ResNet-50) |
| Long-tail Learning | CIFAR-100-LT (ρ=100) | Error Rate | 50.9 | RIDE+distill |
| Long-tail Learning | CIFAR-100-LT (ρ=100) | Error Rate | 52 | RIDE |
| Generalized Few-Shot Learning | ImageNet-LT | Top-1 Accuracy | 56.4 | RIDE (ResNeXt-50) |
| Generalized Few-Shot Learning | ImageNet-LT | Top-1 Accuracy | 54.9 | RIDE (ResNet-50) |
| Generalized Few-Shot Learning | CIFAR-100-LT (ρ=100) | Error Rate | 50.9 | RIDE+distill |
| Generalized Few-Shot Learning | CIFAR-100-LT (ρ=100) | Error Rate | 52 | RIDE |