Few-shot object detection via semantic prompts and classifier decoupling.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot object detection (FSOD) methods often use two-stage detectors for higher accuracy.
• Adapting two-stage detectors to FSOD presents challenges due to limited feature information.

Purpose of the Study:

• To propose a novel FSOD method leveraging semantic prompts and classifier decoupling.
• To effectively fuse and align image and text features for improved object detection.

Main Methods:

• Introduced a Semantic Prompts module to enhance few-shot learning features and image content understanding.
• Employed Gradient Scaling to mitigate negative inter-module influences in two-stage detectors.
• Utilized a Classifier Decoupling module to address inconsistent feature demands between classification and regression branches.

Main Results:

• The proposed method achieved state-of-the-art performance, outperforming baselines like DeFRCN by up to 3.15% mAP under 1-shot settings on PASCAL VOC.
• Demonstrated significant improvements in generalization and robustness within low-data regimes.

Conclusions:

• The novel approach effectively integrates semantic information and decouples classifier branches for superior few-shot object detection.
• The method shows strong potential for real-world applications requiring accurate detection with limited training data.