Meta-DETR: Image-Level Few-Shot Detection With Inter-Class Correlation Exploitation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot object detection commonly uses meta-learning within region-based frameworks.
• Existing methods struggle with low-quality region proposals and ignore inter-class correlations, limiting generalization.

Purpose of the Study:

• To develop an image-level few-shot object detector that addresses limitations of current region-based approaches.
• To introduce a novel meta-learning strategy that captures inter-class correlations for enhanced detection performance.

Main Methods:

• Designed Meta-DETR, the first image-level detector for few-shot object detection.
• Introduced an inter-class correlational meta-learning strategy to leverage relationships between classes.
• Eliminated the need for region proposals, operating solely at the image level.

Main Results:

• Meta-DETR demonstrates superior performance compared to state-of-the-art methods on multiple benchmarks.
• The correlational meta-learning effectively captures inter-class relationships, reducing misclassification.
• Improved generalization of base-class knowledge to novel classes was achieved.

Conclusions:

• Meta-DETR offers a robust and accurate solution for few-shot object detection by operating at the image level.
• The proposed correlational meta-learning strategy significantly enhances detection accuracy and knowledge generalization.
• This approach overcomes key limitations of prior region-based few-shot object detection frameworks.