青年学者论坛-杨洁.pdf

1、1NeUDF:Learning Neural Unsigned Distance Fields with Volume RenderingYu-Tao Liu1,2Li Wang1,2Jie Yang1Weikai Chen3Xiaoxu Meng3Bo Yang3Lin Gao1,2*1Institute of Computing Technology 2University of Chinese Academy of Sciences 3Tencent Games2基于无符号距离场的任意拓扑曲面重建基于无符号距离场的任意拓扑曲面重建闭合开放合成实拍DTU，BMVSNeRF Syntheti

2、cMGNDF3D3基于无符号距离场的任意拓扑曲面重建基于无符号距离场的任意拓扑曲面重建How to represent the complex surfaceSDM-NET Gao et al.2019NeRF Mildenhall et al.2020NeuS Wang et al.2021Gaussian Splatting Kerbl et al.20234基于无符号距离场的任意拓扑曲面重建基于无符号距离场的任意拓扑曲面重建NeRF with UDF reconstruction for open surface reconstructionexpensive,expertise req

3、uiredefficient,user friendlyObjectPhone capturingMulti-view imagesNeural networkLaser scanningPoint cloudManually processingMeshCOLMAP Onberger et al.2016NDF Chibane et al.2020传统方法需要点云媒介NeRF Mildenhall et al.2020NeRF方法直接从图像学习5基于无符号距离场的任意拓扑曲面重建基于无符号距离场的任意拓扑曲面重建NeRF with UDF reconstruction for open su

4、rface reconstructionNeuSSDFInput imagesClosed meshOpen mesh6有符号距离场无符号距离场Venkatesh et al.DUDE:Deep Unsigned Distance Embeddings for Hi-Fidelity Representation of Complex 3D Surfaces输入：多目图像输出：三角网格基于无符号距离场的任意拓扑曲面重建基于无符号距离场的任意拓扑曲面重建Signed Distance v.s.Unsigned Distance7基于无符号距离场的任意拓扑曲面重建基于无符号距离场的任意拓扑曲面重建

5、SDF vs UDFSDF:Learn a criterion to divide the space,easily get stuck in local maximumUDF:Learn local minima,flexible in topologyGTSDFUDF8挑战如何利用体渲染优化神经辐射场（体积表示），进而保证无符号距离场收敛到真实表面（表面表示）图像表示的真值体积表示的神经辐射场UDF体渲染表面表示的三角网格相机采样光线采样点UDF渲染权重颜色损失收敛到表面的神经辐射场自一致无偏遮挡感知网格提取神经辐射场的体积表示三角网格的表面表示基于无符号距离场的任意拓扑曲面重建基于无符号

6、距离场的任意拓扑曲面重建9基于无符号距离场的任意拓扑曲面重建基于无符号距离场的任意拓扑曲面重建Comparison with a naive solutionA naive solution is to directly use the same weight structure of NeuS The naive solution fails to distinguish multiple intersections or passing-by edges,leading to noise andpoorly d