Few-Shot Object Detection: A Comprehensive Survey.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning object detectors typically require extensive annotated data for training.
• This reliance on large datasets poses significant challenges in data acquisition and annotation.

Purpose of the Study:

• To provide a comprehensive overview of the state-of-the-art in few-shot object detection (FSOD).
• To categorize and analyze existing FSOD approaches based on training schemes and architectural designs.
• To identify key concepts for improving performance on novel object categories.

Main Methods:

• Categorization of FSOD approaches by training strategy and network architecture.
• Description of general methodologies and performance-enhancing concepts for each category.
• Review of common datasets, evaluation protocols, and benchmark results.

Main Results:

• Analysis of different FSOD strategies, highlighting effective techniques for novel categories.
• Identification of common challenges and limitations in current FSOD evaluation.
• Summary of promising trends in the rapidly evolving field of FSOD.

Conclusions:

• FSOD is a crucial area for reducing data dependency in object detection.
• Standardized evaluation and further research into novel category adaptation are essential.
• The field shows significant promise for more efficient and adaptable AI vision systems.