Model-Agnostic and Efficient Mixup Augmentation Guided by Saliency Maps.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Mixup-based data augmentation methods leverage saliency information for improved training signals.
• Existing saliency-aware methods often incur high computational costs, require extra modules, or are architecture-specific.

Purpose of the Study:

• To introduce GuidedMixup, a model-agnostic, saliency-aware mixup strategy that overcomes the limitations of prior approaches.
• To develop an efficient algorithm for identifying harmonious image pairs with minimal saliency conflict.

Main Methods:

• GuidedMixup identifies compatible image pairs within mini-batches and uses simplified, fine-grained masks for pixel-wise mixing based on relative saliency.
• GuidedMixup++ incorporates an efficient optimal location search for target image relocation, utilizing convolution operations for rapid conflict assessment.

Main Results:

• GuidedMixup and GuidedMixup++ demonstrate superior efficiency, generalization, and robustness compared to existing saliency-based techniques.
• The proposed methods show significant improvements in downstream tasks, including object detection and instance segmentation.

Conclusions:

• GuidedMixup offers an efficient and effective saliency-aware data augmentation strategy that is model-agnostic.
• The enhanced GuidedMixup++ further improves performance through efficient target image relocation, highlighting its potential for advancing computer vision models.