Surface reconstruction using Signed Distance Fields(SDF)is a prevalent strategy in 3D reconstruction.The study proposes a method that jointly reconstructs SDF with a Trivector grid and a Multilayer Perceptron(MLP),the limitations of achieving a higher resolution and precisely reconstructing surfaces encountered with the existing display grid-based surface reconstruction.The resolution of the Trivector grid exhibits a linear relationship with memory growth,allowing easy scaling to higher levels.This method outperforms singular MLP with enhanced fitting capabilities.The self-attention convolution is used to generate features on different frequency bands,reducing grid discretization and increasing non-linear representation capacity.Moreover,location encoding is embedded into the Trivector feature vectors to combat noise during the surface reconstruction process by introducing inductive bias.Finally,an optimized data sampling method is proposed for dealing with the challenges of fitting complex surfaces,by increasing the sampling frequency around complex surfaces.Experimental results demonstrate that the proposed method surpasses the most advanced ones by 4％in terms of surface reconstruction accuracy on the DTU dataset.
关键词
深度学习/计算机视觉/表面重建
Key words
Deep learning/Computer vision/Surface reconstruction
维普
万方数据