RepCo: Replenish sample views with better consistency for contrastive learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Contrastive learning methods learn representations by comparing sample pairs.
• Current methods often use random image cropping, leading to inconsistent semantic information and suboptimal performance.
• Designing effective sample pairs is crucial for improving contrastive learning.

Purpose of the Study:

• To propose a novel self-supervised learning (SSL) framework, RepCo, to address limitations in sample pair generation for contrastive learning.
• To improve the quality of contrastive views by exploring semantic consistency within a single image.
• To generate more informative representations for downstream tasks.

Main Methods:

• RepCo selects patches from the same image to form positive and negative pairs, unlike previous methods relying on random cropping from different views.
• It encourages similar patches from different positions to be positive pairs.
• It enforces dissimilar patches from adjacent positions to be negative pairs, enriching learned representations.

Main Results:

• RepCo effectively generates high-quality contrastive views by leveraging intra-image semantic consistency.
• The framework provides more informative representations beneficial for various downstream tasks.
• Significant performance gains were observed in object detection (e.g., +2.1 AP50 on Pascal VOC) and semantic segmentation (e.g., +2.3 mIoU on Cityscapes).

Conclusions:

• RepCo offers a novel approach to self-supervised representation learning by intelligently constructing sample pairs from within an image.
• The method overcomes the semantic inconsistency issue associated with random cropping in traditional contrastive learning.
• RepCo demonstrates strong performance improvements on critical computer vision tasks, highlighting its effectiveness.