Efficient Few-Shot Object Detection via Knowledge Inheritance.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot object detection (FSOD) methods have advanced but often neglect efficiency, crucial for embedded AI.
• High computational complexity and slow adaptation hinder practical application of current FSOD techniques.

Purpose of the Study:

• To develop an efficient pretrain-transfer framework (PTF) for FSOD that enhances adaptation speed without computational overhead.
• To introduce novel methods for reliable weight initialization and vector length consistency in FSOD.

Main Methods:

• Proposed an efficient pretrain-transfer framework (PTF) as a baseline.
• Devised a knowledge inheritance (KI) initializer for novel classifier weights.
• Introduced an adaptive length re-scaling (ALR) strategy to manage weight vector inconsistencies.

Main Results:

• Achieved state-of-the-art (SOTA) performance on PASCAL VOC, COCO, and LVIS benchmarks.
• Demonstrated significantly faster adaptation speeds ($1.8-100\times$) on COCO/LVIS compared to existing methods.
• The PTF baseline achieved comparable results to SOTA methods with no computational increment.

Conclusions:

• The developed framework offers a powerful and efficient solution for few-shot object detection.
• This work pioneers the consideration of efficiency in FSOD, aiming to drive future research towards practical, high-performance models.
• The proposed methods facilitate faster knowledge transfer and improve adaptation speed for unseen tasks.