Evaluating Video-to-3D Foundation Models for Wire Harness Assembly Verification

Abstract

Automotive wire harness assembly is a heavily manual process, and a critical part of this workflow is Quality Control (QC). Automating this visual verification is challenging: traditional 2D computer vision struggles with spatial occlusions, and lacks essential geometric depth, whereas 3D data can capture a more comprehensive representation of the object. However, existing 3D machine learning methods depend on rigid scanning hardware and extensive training data. To bypass these technical bottlenecks and reduce the industry’s critical reliance on human labor, this thesis evaluates emerging feed-forward "Video-to-3D" foundation models for automated assembly verification. Using a custom dataset of smartphone videos, four state-of-the-art (SOTA) architectures (Pi3-X, VGGT, Depth Anything V3, and Map Anything) were evaluated for geometric fidelity and automated component classification. Results identify Depth Anything V3 demonstrated superior global spatial coherence and strong baseline counting capabilities, though high false-positive rates across the models indicate further refinement is needed. Pi3-x also proved to be a reliable architecture with competitive classification performance. Although current models lack the geometric precision for highly granular verification tasks, often exhibiting wire duplication or loss of detailed structural definition, they show high promise on broader classification tasks, such as distinguishing and counting clips versus terminal housings. While fully autonomous inspection requires further refinement to mitigate high falsepositive rates, this flexible 3D approach establishes a step toward modernizing visual quality control of wire harnesses