A Semi-Supervised Object Detector Based on Adaptive Weighted Active Learning and Orthogonal Data Augmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Limited labeled data hinders object detection model training.
• Efficient resource utilization is crucial for developing advanced AI models.

Purpose of the Study:

• To propose a semi-supervised object detection (SSOD) approach using adaptive weighted active learning (AWAL) and orthogonal data augmentation (ODA).
• To enhance the utilization of limited annotated data and unlabeled data for improved object detection performance.

Main Methods:

• Implemented an uncertainty sampling framework with adaptively weighted evaluations for informative sample selection.
• Introduced an adaptive weighted loss function to leverage unlabeled data, using normalized uncertainty scores as weights.
• Applied orthogonal data augmentation (ODA) for pseudo-supervised learning on augmented data to capture modality diversity.

Main Results:

• Achieved a mean average precision (mAP) of 35.10 on the MS-COCO dataset using only 10% of annotated data.
• AWAL strategy improved performance by 1.3% compared to baselines without ODA.
• Incorporating ODA yielded an additional 1.2% performance gain.
• Attained 43.30 mAP when trained on fully annotated MS-COCO with additional unlabeled data.

Conclusions:

• The proposed AWAL and ODA approach significantly enhances semi-supervised object detection efficiency and performance.
• The method effectively utilizes limited labeled data and unlabeled data, outperforming existing active learning strategies.
• Demonstrated the approach's superiority and potential for resource-constrained AI development.