Analyzing Target-, Handcrafted- and Learning-Based Methods for Automated 3D Measurement and Modelling

Abstract. In industrial vision metrology, precise spatial measurement is vital for quality control and complex manufacturing, traditionally relying on target arrays for sub-pixel accuracy (0.05–0.1 pixels) and precision to beyond 1:200,000. However, target design, placement and measurement are often time-consuming and challenging for large-scale projects. Automated, markerless methods, generally called Structure-from-Motion (SfM), based on handcrafted algorithms or deep learning-based pipelines, offer greater flexibility but are not widely adopted due not only to concerns about reliability and precision, but also because in many industrial photogrammetry applications targets highlight particular feature points of interest, e.g. tooling points, holes and edges. This study reviews the differences between target-, handcrafted- and learning-based approaches, explores hybrid methods combining targets and natural features, and tests learning-based or handcrafted approaches against the traditional target-based method. Two end-to-end learning-based pipelines based on SuperPoint+LightGlue and KeyNet+AffNet+HardNet are evaluated. Results show that deep learning pipelines for tie point extraction provide enhanced automation but inferior triangulation precision, while being comparable to handcrafted methods.