A Multi-Stage Pipeline for 3D Reconstruction of Vehicles
from Heterogeneous Multi-View Images

Three-dimensional (3D) reconstruction of vehicles remains a challenging task, particularly when aiming to generate standardized models from multiple samples of the same make, model, body type, and year. While traditional Structure-from-Motion approaches excel at reconstructing individual objects, they struggle to create generalized models from diverse multi-view datasets exhibiting variations in lighting, background, and appearance. This study presents a novel automated pipeline for standardized 3D vehicle reconstruction.

The proposed methodology employs a three-stage approach: (1) preprocessing with background removal and adaptive histogram equalization to isolate vehicle regions and normalize illumination; (2) deep learning-based orientation classification using an EfficientNet-B0 architecture to categorize images into eight directional views; and (3) sparse reconstruction via hierarchical localization and COLMAP with a neighbor-based image pairing strategy. Applied to a heterogeneous, multi-source collection, the pipeline successfully generated 3D Gaussian Splatting models for over a hundred distinct vehicle classes — one of the first publicly available datasets of vehicle-class-specific 3D Gaussian Splatting reconstructions derived from heterogeneous multi-source collections.