Jiequan Cui, Zhisheng Zhong, Zhuotao Tian, Shu Liu, Bei Yu, Jiaya Jia
In this paper, we propose the Generalized Parametric Contrastive Learning (GPaCo/PaCo) which works well on both imbalanced and balanced data. Based on theoretical analysis, we observe that supervised contrastive loss tends to bias high-frequency classes and thus increases the difficulty of imbalanced learning. We introduce a set of parametric class-wise learnable centers to rebalance from an optimization perspective. Further, we analyze our GPaCo/PaCo loss under a balanced setting. Our analysis demonstrates that GPaCo/PaCo can adaptively enhance the intensity of pushing samples of the same class close as more samples are pulled together with their corresponding centers and benefit hard example learning. Experiments on long-tailed benchmarks manifest the new state-of-the-art for long-tailed recognition. On full ImageNet, models from CNNs to vision transformers trained with GPaCo loss show better generalization performance and stronger robustness compared with MAE models. Moreover, GPaCo can be applied to the semantic segmentation task and obvious improvements are observed on the 4 most popular benchmarks. Our code is available at https://github.com/dvlab-research/Parametric-Contrastive-Learning.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Domain Adaptation | ImageNet-R | Top-1 Error Rate | 39.7 | GPaCo (ViT-L) |
| Domain Adaptation | ImageNet-C | mean Corruption Error (mCE) | 39 | GPaCo (ViT-L) |
| Domain Adaptation | ImageNet-Sketch | Top-1 accuracy | 48.3 | GPaCo (ViT-L) |
| Semantic Segmentation | PASCAL Context | mIoU | 56.2 | GPaCo (ResNet101) |
| Semantic Segmentation | ADE20K | Validation mIoU | 54.3 | GPaCo (Swin-L) |
| Image Classification | Places-LT | Top-1 Accuracy | 41.7 | GPaCo (ResNet-152) |
| Image Classification | ImageNet-LT | Top-1 Accuracy | 63.2 | GPaCo (2-ResNeXt101-32x4d) |
| Few-Shot Image Classification | Places-LT | Top-1 Accuracy | 41.7 | GPaCo (ResNet-152) |
| Few-Shot Image Classification | ImageNet-LT | Top-1 Accuracy | 63.2 | GPaCo (2-ResNeXt101-32x4d) |
| Generalized Few-Shot Classification | Places-LT | Top-1 Accuracy | 41.7 | GPaCo (ResNet-152) |
| Generalized Few-Shot Classification | ImageNet-LT | Top-1 Accuracy | 63.2 | GPaCo (2-ResNeXt101-32x4d) |
| Long-tail Learning | Places-LT | Top-1 Accuracy | 41.7 | GPaCo (ResNet-152) |
| Long-tail Learning | ImageNet-LT | Top-1 Accuracy | 63.2 | GPaCo (2-ResNeXt101-32x4d) |
| Generalized Few-Shot Learning | Places-LT | Top-1 Accuracy | 41.7 | GPaCo (ResNet-152) |
| Generalized Few-Shot Learning | ImageNet-LT | Top-1 Accuracy | 63.2 | GPaCo (2-ResNeXt101-32x4d) |
| Domain Generalization | ImageNet-R | Top-1 Error Rate | 39.7 | GPaCo (ViT-L) |
| Domain Generalization | ImageNet-C | mean Corruption Error (mCE) | 39 | GPaCo (ViT-L) |
| Domain Generalization | ImageNet-Sketch | Top-1 accuracy | 48.3 | GPaCo (ViT-L) |
| 10-shot image generation | PASCAL Context | mIoU | 56.2 | GPaCo (ResNet101) |
| 10-shot image generation | ADE20K | Validation mIoU | 54.3 | GPaCo (Swin-L) |