Accepted: IJCV 2024 | Yanbiao Ma, Licheng Jiao, Fang Liu, Shuyuan Yang, Xu Liu and Puhua Chen
This paper addresses the challenge of long-tailed classification, where the majority of samples belong to a few classes (head classes), while the minority samples are distributed across many other classes (tail classes). The authors propose a method that leverages geometric prior knowledge to guide the model in learning the underlying distribution of tail classes. The key contributions are:
1. **Geometric Prior Guided Feature Representation (FUR):** The method defines the geometry of feature distributions and introduces a similarity measure between these geometries. Four experimental phenomena are discovered, showing that similar feature distributions have similar geometries, and the similarity decreases as inter-class similarity decreases.
2. **Feature Uncertainty Representation:** The proposed method models the uncertainty representation of tail features using the geometry of head class features. This involves perturbing tail features to cover the underlying distribution of the tail class, allowing the model to learn valuable information outside the observed domain.
3. **Three-Stage Training Scheme:** A three-stage training scheme is designed to apply the feature uncertainty representation effectively. The first two stages use decoupled training, and the third stage fine-tunes the feature extractor to adapt to the new decision boundaries, improving overall model performance.
4. **Experimental Validation:** The method is evaluated on several benchmark datasets, including CIFAR-10/100-LT, ImageNet-LT, iNaturalist 2018, and OIA-ODIR. The results show that the proposed method outperforms other state-of-the-art methods, particularly in improving the performance of tail classes.
The paper provides a comprehensive theoretical foundation and practical approach to address the long-tailed classification challenge, offering new insights and methods for improving model generalization on imbalanced datasets.This paper addresses the challenge of long-tailed classification, where the majority of samples belong to a few classes (head classes), while the minority samples are distributed across many other classes (tail classes). The authors propose a method that leverages geometric prior knowledge to guide the model in learning the underlying distribution of tail classes. The key contributions are:
1. **Geometric Prior Guided Feature Representation (FUR):** The method defines the geometry of feature distributions and introduces a similarity measure between these geometries. Four experimental phenomena are discovered, showing that similar feature distributions have similar geometries, and the similarity decreases as inter-class similarity decreases.
2. **Feature Uncertainty Representation:** The proposed method models the uncertainty representation of tail features using the geometry of head class features. This involves perturbing tail features to cover the underlying distribution of the tail class, allowing the model to learn valuable information outside the observed domain.
3. **Three-Stage Training Scheme:** A three-stage training scheme is designed to apply the feature uncertainty representation effectively. The first two stages use decoupled training, and the third stage fine-tunes the feature extractor to adapt to the new decision boundaries, improving overall model performance.
4. **Experimental Validation:** The method is evaluated on several benchmark datasets, including CIFAR-10/100-LT, ImageNet-LT, iNaturalist 2018, and OIA-ODIR. The results show that the proposed method outperforms other state-of-the-art methods, particularly in improving the performance of tail classes.
The paper provides a comprehensive theoretical foundation and practical approach to address the long-tailed classification challenge, offering new insights and methods for improving model generalization on imbalanced datasets.