Smart Sampling for Data-Efficient Three-Dimensional Reconstruction of Additively Manufactured Parts

Abstract

Additive manufacturing can produce complex three-dimensional (3D) geometries that can be challenging to inspect and measure. One potential solution is photogrammetry, which can reconstruct 3D shapes from 2D photographs. However, photogrammetry requires many photographs covering the entire surface of the object being inspected, introducing time and cost that are barriers to real-world application. This paper develops a novel method to reduce the number of photographs required for accurate 3D reconstruction. The method automatically selects views showing intricate features of the part and minimizes redundancy for views covering less complex regions. The image selection is based on an image difference score that is derived from a feature matching algorithm based on deep learning (LightGlue). The smart-sampled image set significantly enhances 3D reconstruction fidelity compared to an identical number of equally spaced images, achieving an order of magnitude reduction in Mean Absolute Error (MAE) and a 10-20 dB improvement in Peak Signal-to-Noise Ratio (PSNR). Notably, a smart-sampled set of just 15 images provides comparable accuracy to 60 equally spaced images, and achieves a 33% reduction in compute time.