CAD-free fine-grained industrial part recognition via 3D Gaussian Splatting

Abstract Fine-grained industrial parts such as screw, nut, and gear variants differ only in subtle threading, tooth count, or dimensions, and CAD models and large labeled datasets are rarely available in production. We present a CAD-free, video-based recognizer that distinguishes near-identical parts from a single short turntable video per part, with no CAD or manual labels. From the video we reconstruct a 3D Gaussian Splatting (3DGS) twin of each part and render it from 360 viewpoints as a label-free training source. A compact two-layer head, meta-trained over a frozen DINOv2 backbone with a 1-nearest-neighbor (1-NN)-consistent objective, classifies a real part by its most similar reference view. On a self-collected 18-part dataset, evaluated on a separately captured test video from the same rig to control leakage, it reaches 91.6% accuracy from 10 real shots and 75.5% with zero real labels; meta-learning and a strong frozen backbone are both decisive. We compare the renders with the video’s own masked frames, the CAD-free alternative that needs no reconstruction. On the same rig the frames are the stronger source, by 2.4 points at 10 shots and 6.7 points zero-shot with matched view counts, because fine thread relief is not reproduced faithfully. In a hand-held phone capture under different lighting, surface, and viewpoints the order reverses: the renders lead by 11.7 points zero-shot with matched views and by 14 points on nuts, and labeled shots taken on the rig do not transfer. We also report the twin’s reconstruction fidelity, blind zones, and computational cost.