Core Concepts
MSDA methods exhibit class dependency, mitigated by DropMix technique.
Abstract
The paper explores the class dependency of Mixed Sample Data Augmentation (MSDA) methods like Mixup, CutMix, and PuzzleMix. It introduces the DropMix technique to reduce class dependency and improve classification efficiency. The study includes experiments on CIFAR-100 and ImageNet datasets, revealing the impact of MSDA on different classes and the effectiveness of DropMix in mitigating class dependency.
Introduction
MSDA techniques like Mixup, CutMix, and PuzzleMix enhance performance but exhibit class dependency.
Experiments
Tested MSDA methods on CIFAR-100 and ImageNet datasets.
Introduced evaluation metrics for class dependency.
Results
MSDA methods show class dependency, mitigated by DropMix.
Discussion
Analyzed reasons for class dependency and the impact of MSDA on label information.
Open Problems
Proposed hypotheses to explain variations in class dependency among MSDA methods.
Conclusion
Proposed DropMix technique effectively reduces class dependency in MSDA.
Stats
Mixup improves the average accuracy and recall for the "Lamp" class but reduces recall for the "Dolphin" class.
MSDA methods exhibit class dependency, mitigated by DropMix technique.
Quotes
"MSDA methods exhibit class dependency, mitigated by DropMix technique."