Neural Implicit Surface Reconstruction from Noisy Camera Observations Sarthak Gupta1Patrik Huber2 1Indian Institute of Technology Roorkee Roorkee Uttarakhand India - 247667

2025-05-02 0 0 1.89MB 4 页 10玖币

侵权投诉

Neural Implicit Surface Reconstruction from Noisy Camera Observations

Sarthak Gupta,1Patrik Huber, 2

1Indian Institute of Technology Roorkee, Roorkee, Uttarakhand, India - 247667

2University of York, Deramore Lane, Heslington, York, YO10 5GH, United Kingdom

mrsarthakgupta@gmail.com, patrik.huber@york.ac.uk

Abstract

Representing 3D objects and scenes with neural radiance

ﬁelds has become very popular over the last years. Recently,

surface-based representations have been proposed, that allow

to reconstruct 3D objects from simple photographs. However,

most current techniques require an accurate camera calibra-

tion, i.e. camera parameters corresponding to each image,

which is often a difﬁcult task to do in real-life situations. To

this end, we propose a method for learning 3D surfaces from

noisy camera parameters. We show that we can learn camera

parameters together with learning the surface representation,

and demonstrate good quality 3D surface reconstruction even

with noisy camera observations.

Introduction

The idea of representing 3D objects and scenes with neural

networks, instead of traditional mesh-like representations,

has gained signiﬁcant traction recently. In (Mildenhall et al.

2020), an approach called NeRF is proposed, Neural Radi-

ance Fields, where a neural network is used together with

a volumetric representation to learn the representation of a

scene from a collection of calibrated multi-view 2D images.

However, a volumetric representation is not the best rep-

resentation in many cases; for example, many objects like

faces are better represented using surfaces. This was tackled

in a follow-up work by (Wang et al. 2021a), called NeuS,

who proposed to use neural implicit surfaces together with

volume rendering for multi-view reconstruction. This year,

(Wang et al. 2021b) addressed another shortcoming of clas-

sical NeRF’s, by extending the NeRF method to work on im-

age data when camera calibration data is not present. This is

achieved by jointly estimating the scene representation and

optimising for the camera parameters, and the authors have

shown promising results for frontal-facing scenes.

In this work, we propose to marry the beneﬁts of each

of these approaches: We propose a method to learn a neu-

ral implicit surface based representation of objects from

noisy camera observations. We show that the classical NeuS

method fails to learn an object completely if camera param-

eters are not precise, whereas our approach succeeds. Fur-

thermore, what is often taken as ground-truth camera pa-

rameters come themselves from a multi-view reconstruction

Figure 1: (Left): Reconstruction using (Wang et al. 2021a)

with ground truth camera parameters. (Centre): Reconstruc-

tion using (Wang et al. 2021a) with noisy camera parame-

ters, where the approach completely fails. (Right): Recon-

struction using the proposed approach with noisy camera

parameters. For each of the two objects, the lower images

represent the actual image of the object while the upper im-

age is the rendered image of the reconstructed surface.

software like COLMAP (Sch¨

onberger and Frahm 2016), and

thus likely with estimation errors. By making camera param-

eters learnable, the approach gains the ability to correct for

these errors, and to potentially achieve better quality than

methods that take camera parameters as given truth.

Methodology

We draw inspiration from (Wang et al. 2021b), where the

authors make the camera parameters of a NeRF learnable

to allow them to converge to values that result in desirable

3D scene reconstructions. While (Wang et al. 2021b) works

with unknown camera parameters for only forward-facing

input images with rotational and translational perturbations

of up to ±20◦, our approach works successfully on images

from 360◦view angles. The latter results in a problem that

is much harder to solve and is prone to local optima that do

not produce good results when trying to learn the camera

arXiv:2210.01548v1 [cs.CV] 2 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

NeuralImplicitSurfaceReconstructionfromNoisyCameraObservationsSarthakGupta,1PatrikHuber,21IndianInstituteofTechnologyRoorkee,Roorkee,Uttarakhand,India-2476672UniversityofYork,DeramoreLane,Heslington,York,YO105GH,UnitedKingdommrsarthakgupta@gmail.com,patrik.huber@york.ac.ukAbstractRepresenting3Dobjec...

展开>> 收起<<

Neural Implicit Surface Reconstruction from Noisy Camera Observations Sarthak Gupta1Patrik Huber2 1Indian Institute of Technology Roorkee Roorkee Uttarakhand India - 247667.pdf

共4页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Neural Implicit Surface Reconstruction from Noisy Camera Observations Sarthak Gupta1Patrik Huber2 1Indian Institute of Technology Roorkee Roorkee Uttarakhand India - 247667

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: