Towards Efficient 3D Reconstruction from UAV Imagery: Evaluation of Feature Detection, Description and Matching Combinations

Abstract

This thesis develops a comprehensive efficiency evaluation framework for 3D reconstruction from UAV imagery, addressing the critical need for objective algorithm selection and operational optimization. The research establishes systematic methodologies for evaluating computational efficiency, reconstruction quality, and cross-platform performance through multi-dimensional analysis rather than single-metric assessments. A novel Composite Unsupervised Efficiency Score (CUES) methodology enables multi-dimensional evaluation by integrating four distinct weighting approaches for objective algorithm assessment. The developed framework systematically evaluates 784 feature detection, description, and matching combinations across diverse datasets, providing evidence-based recommendations for algorithm selection, parameter optimization, and timing-based performance metrics. Statistical validation demonstrates consistent methodology reliability with high cross-dataset stability, strong ranking consistency between computing environments, and 95% confidence interval quantification enabling significance testing between algorithm alternatives. The framework includes enhanced COLMAP integration, interactive visualization tools, comprehensive performance databases, and standardized evaluation protocols. Key contributions include parameter optimization guidelines, cross-platform benchmarking across server and mobile architectures, multi-layered synthetic UAV datasets development, and novel efficiency metrics for operational UAV scenarios. This research establishes the first systematic efficiency evaluation framework for UAV-based 3D reconstruction, providing practical guidance for operational deployment optimization and setting new standards for objective performance assessment in computer vision applications.