Expert teacher based on foundation image segmentation model for object detection in aerial images

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection in aerial imagery is hindered by a scarcity of labeled data, limiting model robustness and generalization.
• Existing teacher-student learning methods are primarily explored in natural image domains, with limited focus on unlabeled aerial datasets.

Purpose of the Study:

• To propose an expert teacher framework based on a foundation image segmentation model (ET-FSM) for enhancing object detection in unlabeled aerial images.
• To leverage foundation models for generating high-quality pseudo-labels to improve the performance of student detectors.

Main Methods:

• Developed an expert teacher framework (ET-FSM) utilizing a foundation image segmentation model.
• Incorporated an expert guidance mechanism within a binary detector to leverage knowledge from the foundation model for accurate object localization.
• Introduced a momentum contrast classification module to differentiate between visually similar object categories in aerial imagery.

Main Results:

• The proposed ET-FSM framework effectively generates high-quality pseudo-labels for unlabeled aerial images.
• Experiments demonstrated superior performance compared to existing methods across various student detectors.
• Achieved fully supervised performance levels with significantly reduced amounts of labeled aerial data.

Conclusions:

• The ET-FSM framework offers a viable solution for improving object detection in aerial imagery using unlabeled data.
• Foundation models can be effectively adapted to provide expert guidance for object detection tasks.
• The approach significantly reduces the need for extensive labeled aerial datasets, enhancing efficiency and accessibility.