基于语义嵌入三维高斯溅射的双路径重建算法

打开文本图片集

关键词：三维高斯溅射；语义特征投影；双路径协同渲染；曝光补偿

中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2026）04-034-1245-06

doi：10. 19734/j.issn. 1001-3695.2025.06.0263

Dual-path reconstruction algorithm based on semantically embedded 3D Gaussian splatting

WangBo，Li Shaobo†，Pan Jiaxing，Ren Hao（SchlofAutomationdElectricalEnginering，erogoliaUniersityfiecendchgBotouNeiongolOin）

Abstract：3DGSlacksthecapabilitytoincorporatesemanticinformationwithinitsexplicit Gausiandistributions，whichrestricts itsuseinsemanticperceptionreconstruction.Inresponse totheabove，thiswork proposedadual-pathreconstructionalgorithmthatembeddedsemanticsinto3DGaussansplating.Firstly，thealgorithmoperatedbyprojecting2Dsemanticfeatures frompre-trainedmodelsinto3DGaussiandistributions，utilizing theirspatialrelationships.Then，adual-pathcolaborativerendering strategyfused fastrasterizedrenderinganddepth-awarevolumerendering toenhance output quality.Finaly，anexposure-compensated image lossoptimizationmodelsuppressdartifactscausedbylightingvariations.Experimentalresultsothe Tanks&Tempels dataset show that SDP-GS improves PSNR and SSIM by 9% and 8% ，respectively，and reduces LPIPS by 41% （204号 compared to the baseline PSNR and SSIM.SDP-GS notonlysuccessfullyembeds the semantic features into the 3DGaussian splatting model，but also realizes the high-fidelity reconstruction.

Keywords：3DGaussiansplating；semantic feature projection；dual-pathcollaborativerendering；exposurecompensation

0 引言

三维计算机视觉的核心任务之一是高效、精准地重建三维场景，广泛应用于自动驾驶、增强现实和医疗导航等多个领域。（剩余18321字）