S³H: Long-tailed classification via spatial constraint sampling, scalable network, and hybrid task.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Long-tailed classification presents challenges in creating clear decision boundaries due to imbalanced data distributions.
• Existing methods struggle to effectively integrate semantic and textural information for robust feature extraction.

Purpose of the Study:

• To develop an end-to-end trainable method for long-tailed classification that improves feature representation.
• To enhance the model's ability to capture well-balanced features by integrating semantic consistency and texture characteristics.

Main Methods:

• Proposed a spatial constraint sampling strategy to provide models with representative features.
• Introduced a scalable network architecture for dynamic feature adjustment.
• Developed a hybrid task integrating single-model classification and cross-model contrastive learning for comprehensive feature capture.

Main Results:

• Achieved state-of-the-art performance on CIFAR10-LT, CIFAR100-LT, ImageNet-LT, and iNaturalist 2018 datasets.
• Demonstrated effective integration of high-level semantic and low-level texture features.
• Enabled end-to-end training, overcoming multi-stage optimization constraints.

Conclusions:

• The proposed method offers a novel and effective solution for long-tailed classification.
• The combination of spatial constraint sampling, scalable networks, and hybrid tasks leads to superior feature learning.
• This approach paves the way for more robust and efficient long-tailed classification models.