结合平面约束与Kolmogorov-Arnold网络的室内场景三维重建方法

打开文本图片集

关键词：室内三维重建；神经辐射场；平面约束优化模块；KAN颜色分支模块中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2025）10-035-3167-07doi：10.19734/j.issn.1001-3695.2025.01.0009

Indoor scene 3D reconstruction method combining plane constraints and Kolmogorov-Arnold network

Jia Di 1，2 ，Meng Xiaohual†，Cheng Shuo1，Xu Chi 1 ，Liu Yang1，Song Huilun1 （1.OrdosIstituteofLiingTechicalUniersit，OrdosNeiMongolOooChina;2.SchlofEletroic &IforationEnging， LiaoningTechnical University，HuludaoLiaoning1251o5，China）

Abstract：Toaddressthechallnges ofweak texture areas in indoor3Dreconstructionand the limitedabilityto express complex scenes，thispaper proposedanindoor3Dreconstruction network，KAN-PlaneNet，which combined aplaneconstraintoptimizationmoduleandaKANcolorbranch module.Basedonneuralradiancefields（NeRF），thenetwork introducedthe KAN moduleandcoloratributeinputs inthecolorbranchtoenhancethedetail expressionofcomplexscenes.Inaddition，byintroducinga planeconstraintoptimization module，itenforcedverticalalignmentbetweenwallandflorregions.Itdefinedalearnable wall normal vector nw to constrain the normal direction of surface points on the wall to be parallel or orthogonal to nw .For indoorobjectregions，tappliedaconstraintrequiringthecosinesimilaritybetweenthenormaldirectionoftheregionandthe floornormaltobegreaterthanO.95，thusenforcing paralelism.Experimentalresultsshowthatthemethodoutperformsexisting 3Dsupervised methodsontheScanNet dataset.ComparedtotheVolSDFmethod，theoverall evaluation metric improves by an average of O.356，validating its superiority.

Key words：indoor3Dreconstruction；neural radiance fields；plane constraint optimization module；KANcolor branch module

0 引言

室内三维重建技术对现代社会产生了深远影响，特别是在建筑、城市规划和智能家居等领域。（剩余17391字）