Decoupled Metric Network for Single-Stage Few-Shot Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot object detection is a challenging problem despite advances in general object detection.
• Existing methods struggle with limited training data, hindering performance.

Purpose of the Study:

• To propose a novel decoupled metric network (DMNet) for single-stage few-shot object detection.
• To address the challenges of limited data and representation disagreement in few-shot detection.

Main Methods:

• Introduced Decoupled Representation Transformation (DRT) to predict objectness and anchor shape, mitigating handcrafted prior effects.
• Developed Image-Level Distance Metric Learning (IDML) for enhanced generalization in few-shot classification within detection.
• Designed DMNet as a single-stage metric detection paradigm, integrating DRT and IDML.

Main Results:

• DMNet achieves state-of-the-art performance on PASCAL VOC and MS COCO datasets.
• The decoupled approach effectively handles representation disagreement for better learning from limited data.
• IDML improves generalization by performing metric learning on entire visual features, enabling efficient parallel inference.

Conclusions:

• The proposed DMNet offers an effective single-stage metric detection approach for few-shot learning.
• DRT and IDML are key components enabling robust few-shot object detection.
• The method demonstrates significant improvements in few-shot object detection tasks.