Zhengzhuo Xu, Ruikang Liu, Shuo Yang, Zenghao Chai, Chun Yuan
The real-world data tends to be heavily imbalanced and severely skew the data-driven deep neural networks, which makes Long-Tailed Recognition (LTR) a massive challenging task. Existing LTR methods seldom train Vision Transformers (ViTs) with Long-Tailed (LT) data, while the off-the-shelf pretrain weight of ViTs always leads to unfair comparisons. In this paper, we systematically investigate the ViTs' performance in LTR and propose LiVT to train ViTs from scratch only with LT data. With the observation that ViTs suffer more severe LTR problems, we conduct Masked Generative Pretraining (MGP) to learn generalized features. With ample and solid evidence, we show that MGP is more robust than supervised manners. In addition, Binary Cross Entropy (BCE) loss, which shows conspicuous performance with ViTs, encounters predicaments in LTR. We further propose the balanced BCE to ameliorate it with strong theoretical groundings. Specially, we derive the unbiased extension of Sigmoid and compensate extra logit margins to deploy it. Our Bal-BCE contributes to the quick convergence of ViTs in just a few epochs. Extensive experiments demonstrate that with MGP and Bal-BCE, LiVT successfully trains ViTs well without any additional data and outperforms comparable state-of-the-art methods significantly, e.g., our ViT-B achieves 81.0% Top-1 accuracy in iNaturalist 2018 without bells and whistles. Code is available at https://github.com/XuZhengzhuo/LiVT.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 8.7 | ViT-B + Bal-BCE |
| Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 9.3 | ViT-B + Bal-CE |
| Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 10.1 | ViT-B + CB |
| Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 10.5 | ViT-B + CE |
| Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 11.4 | ViT-B + LDAM |
| Few-Shot Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 8.7 | ViT-B + Bal-BCE |
| Few-Shot Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 9.3 | ViT-B + Bal-CE |
| Few-Shot Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 10.1 | ViT-B + CB |
| Few-Shot Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 10.5 | ViT-B + CE |
| Few-Shot Image Classification | CIFAR-10-LT (ρ=10) | Error Rate | 11.4 | ViT-B + LDAM |
| Generalized Few-Shot Classification | CIFAR-10-LT (ρ=10) | Error Rate | 8.7 | ViT-B + Bal-BCE |
| Generalized Few-Shot Classification | CIFAR-10-LT (ρ=10) | Error Rate | 9.3 | ViT-B + Bal-CE |
| Generalized Few-Shot Classification | CIFAR-10-LT (ρ=10) | Error Rate | 10.1 | ViT-B + CB |
| Generalized Few-Shot Classification | CIFAR-10-LT (ρ=10) | Error Rate | 10.5 | ViT-B + CE |
| Generalized Few-Shot Classification | CIFAR-10-LT (ρ=10) | Error Rate | 11.4 | ViT-B + LDAM |
| Long-tail Learning | CIFAR-10-LT (ρ=10) | Error Rate | 8.7 | ViT-B + Bal-BCE |
| Long-tail Learning | CIFAR-10-LT (ρ=10) | Error Rate | 9.3 | ViT-B + Bal-CE |
| Long-tail Learning | CIFAR-10-LT (ρ=10) | Error Rate | 10.1 | ViT-B + CB |
| Long-tail Learning | CIFAR-10-LT (ρ=10) | Error Rate | 10.5 | ViT-B + CE |
| Long-tail Learning | CIFAR-10-LT (ρ=10) | Error Rate | 11.4 | ViT-B + LDAM |
| Generalized Few-Shot Learning | CIFAR-10-LT (ρ=10) | Error Rate | 8.7 | ViT-B + Bal-BCE |
| Generalized Few-Shot Learning | CIFAR-10-LT (ρ=10) | Error Rate | 9.3 | ViT-B + Bal-CE |
| Generalized Few-Shot Learning | CIFAR-10-LT (ρ=10) | Error Rate | 10.1 | ViT-B + CB |
| Generalized Few-Shot Learning | CIFAR-10-LT (ρ=10) | Error Rate | 10.5 | ViT-B + CE |
| Generalized Few-Shot Learning | CIFAR-10-LT (ρ=10) | Error Rate | 11.4 | ViT-B + LDAM |