ECEA: Extensible Co-Existing Attention for Few-Shot Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot object detection (FSOD) methods often focus on global features, neglecting local-to-global object localization.
• Limited training data in FSOD leads to incomplete object samples, hindering the detection of unseen objects.

Purpose of the Study:

• To propose an Extensible Co-Existing Attention (ECEA) module for FSOD that infers global objects from local parts.
• To enhance the ability of FSOD models to detect objects even when training samples capture only partial views.

Main Methods:

• Devised an extensible attention mechanism to expand from local regions to similar/adjacent co-existing regions.
• Implemented extensible attention across multiple feature scales for progressive object discovery.
• Utilized a two-stage learning paradigm: base stage for learning extensibility, novel stage for adaptation.

Main Results:

• The ECEA module enables complete object prediction despite missing training regions.
• Achieved new state-of-the-art performance on PASCAL VOC and COCO datasets.
• Demonstrated effective transfer of extensible learning from base to novel classes.

Conclusions:

• The ECEA module significantly improves FSOD performance by addressing partial object detection.
• The proposed method offers a robust solution for detecting objects with scarce annotated data.
• Code availability facilitates further research and application of the ECEA module.