Ensemble-Based Student-Teacher Refinement Framework for Semi-Supervised Object Detection (ESTR)

Semi-supervised object detection (SSOD) reduces the need for costly manual annotation by combining limited labeled data with larger collections of unlabeled images. However, existing approaches often rely on a single teacher model or iterative teacher–student training, which can introduce noisy pseudo-labels and increase computational complexity. They also typically require the student to share the same architecture as the teacher, limiting flexibility.

This thesis proposes Ensemble-Based Student-Teacher Refinement (ESTR), an ensemble-based framework that generates more reliable pseudo-labels through agreement among multiple teacher models. The main contribution is a consensus-based refinement algorithm that uses cross-class IoU clustering and majority voting to enforce spatial and semantic consistency. Low-consensus predictions are discarded, while agreed-upon boxes are averaged to produce robust pseudo-labels.

ESTR further introduces a decoupled teacher–student architecture, separating pseudo-label generation stage from student training stage. This allows the student to adopt a different detection architecture from the teacher ensemble.

Evaluated on the Pascal VOC benchmark, ESTR achieves 80.22% AP50 and 59.9% AP50:95, demonstrating improved pseudo-label reliability while providing a flexible and architecture-agnostic SSOD framework.

Supervisor:  Dr. Abbas A. Fairouz
Convener:  Dr. Jassim M. Aljuraidan
Examination Committee: Prof. Maytham Safar