Experimental Evaluation of an Image-Based 3D Reconstruction Pipeline Using Coherent Point Drift Alignment

Abstract

This paper presents an experimental evaluation of a modular Python pipeline for image-based 3D reconstruction and quantitative point cloud comparison. The pipeline integrates learned feature extraction and matching (SuperPoint, NetVLAD, LightGlue) with classical Structure-from-Motion and Multi-View Stereo reconstruction (COLMAP), followed by Interquartile-Range-based scale estimation and Coherent Point Drift (CPD) alignment against a ground truth reference. Reconstruction quality is assessed using six complementary metrics: Chamfer Distance, Hausdorff Distance, Root Mean Square Error, Completeness, Accuracy, and F-Score. The pipeline is validated on two datasets from the Tohoku University Multi-View Stereo (THU-MVS) benchmark, comprising photographs of a painted cat figurine (108 images) and a painted dog figurine (72 images), each paired with a structured-light ground truth scan. After manual removal of background geometry from the reconstructed clouds, the proposed pipeline achieves F-Scores of 0.936 and 0.984 at a strict threshold of $\tau=5 \mathrm{~mm}$, reaching 99.93% and ${1 0 0 \%}$ at $\tau=20 \mathrm{~mm}$, with Chamfer Distances of 3.249 and 2.252 scene units. These results demonstrate that the combination of learned feature matching and probabilistic CPD alignment yields highly accurate, complete point cloud reconstructions on richly textured objects, while highlighting background removal as a critical preprocessing step for reliable scale estimation and registration.