CAD models clustering with machine learning

Abstract

Similarity assessment between 3D models is an important problem in many fields including medicine, biology and industry. As there is no direct method to compare 3D geometries, different model representations (shape signatures) are developed to enable shape description, indexing and clustering. Even though some of those descriptors proved to achieve high classification precision, their application is often limited. In this work, a different approach to similarity assessment of 3D CAD models was presented.Insteadoffocusingononespecificshapesignature,45easy-to-extractshape signatureswereconsideredsimultaneously.Thevectorofthosefeaturesconstituted aninputfor3machinelearningalgorithms:therandomforestclassifier,thesupport vectorclassifierandthefullyconnectedneuralnetwork.Theusefulnessoftheproposed approachwasevaluatedwithadatasetconsistingofover1600CADmodelsbelonging to9separateclasses.Differentvaluesofhyperparameters,aswellasneuralnetwork configurations,wereconsidered.Retrievalaccuracyexceeding99%wasachievedon thetestdataset.