LGCOAMix: Local and Global Context-and-Object-Part-Aware Superpixel-Based Data Augmentation for Deep Visual...

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Cutmix data augmentation improves generalization but overlooks local context, hindering performance.
• Existing methods lose object part information and require inefficient processing for label consistency.

Purpose of the Study:

• To introduce LGCOAMix, an efficient data augmentation method addressing limitations of current Cutmix approaches.
• To improve attention to discriminative local features and object parts in deep learning models.

Main Methods:

• LGCOAMix employs a superpixel-based grid blending strategy for context-aware and object-part-aware data augmentation.
• Utilizes a novel superpixel attention approach for label mixing in Cutmix-based augmentation.
• Learns local features from discriminative superpixel regions and cross-image superpixel contrasts.

Main Results:

• LGCOAMix significantly outperforms state-of-the-art Cutmix methods on classification tasks.
• Demonstrates superior performance in weakly supervised object localization on the CUB200-2011 dataset.
• Proves effective for both Convolutional Neural Networks (CNNs) and Transformer networks.

Conclusions:

• LGCOAMix offers an efficient and effective approach to data augmentation, enhancing model generalization.
• The superpixel attention mechanism is a key innovation for improving Cutmix-based methods.
• LGCOAMix provides a versatile solution applicable to various deep learning architectures and tasks.