Enhanced Long-Tailed Recognition With Contrastive CutMix Augmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Real-world datasets often exhibit long-tailed distributions, with few dominant classes and numerous rare classes.
• Deep learning models struggle with generalization on tail classes due to data imbalance, leading to poor performance.

Purpose of the Study:

• To address the limitations of traditional CutMix in handling imbalanced data by incorporating semantic information.
• To propose Contrastive CutMix (ConCutMix) for improved long-tailed recognition.

Main Methods:

• Contrastive learning is employed to derive semantic similarities between image samples.
• These semantic similarities are used to refine the area-based labels generated by CutMix, creating semantically consistent labels.
• The proposed ConCutMix method augments data for tail classes to enhance model generalization.

Main Results:

• ConCutMix significantly enhances accuracy on tail classes and overall performance in long-tailed recognition tasks.
• Experiments on ImageNet-LT using ResNeXt-50 demonstrated a 3.0% overall accuracy improvement, with a 3.3% gain on tail classes.
• The effectiveness of ConCutMix was validated across different benchmarks and model architectures.

Conclusions:

• Contrastive CutMix offers a robust solution for improving deep learning performance on imbalanced datasets.
• The method's ability to generate semantically consistent labels is key to its success in long-tailed recognition.
• ConCutMix provides a valuable technique for enhancing model generalization in real-world scenarios with data imbalance.