A foundational architecture prioritizing transparency and low computational overhead over opaque deep learning methods. Built for industrial repositories containing millions of 3D models.
Industrial repositories contain millions of 3D models, but engineers constantly recreate existing parts because text-based metadata searches rely on inconsistent human tagging. Deep learning alternatives are often unviable because they require expensive GPU hardware, extensive labeled data, and act as black boxes.
A training-free, CPU-only retrieval framework that allows engineers to search for components based directly on their physical shape and volume, enabling rapid design reuse without the overhead of heavy neural networks.
The system features a mathematical descriptor contribution analysis that decomposes every match score into interpretable components. This provides transparent evidence for why a part matched, overcoming the trust barriers associated with deep learning models.
Evaluated on an industrial dataset of 600 mechanical models across 15 distinct classes, the engine hits a Precision@5 of 91.50% and a Precision@10 of 83.25% with sub-millisecond query times.
Raw 3D meshes are translated to the origin, scaled uniformly so their largest bounding box dimension equals one, and oriented using Principal Component Analysis to guarantee a consistent coordinate frame across the repository.
The aligned model is orthogonally sliced 100 times along its X, Y, and Z axes. These binary cross-sections are accumulated into 2D density maps that accurately capture internal spatial volume rather than just boundary silhouettes.
The system extracts a Histogram of Oriented Gradients (HOG) from the MDOR maps to encode fine local edge patterns. This is complemented by two statistical shape descriptors: D1 (capturing radial mass distribution from the centroid) and D2 (capturing pairwise surface-point distances to determine global compactness).
The three descriptors are Z-score standardized and fused into a 5420-dimensional vector. PCA compression reduces this down to just 86 components while retaining 95% of the original variance.