1 of 34

NeRO: Neural Geometry and BRDF Reconstruction of Reflective Objects from Multiview Images

Liu et. al.

2 of 34

TOPICS OF DISCUSSION

Introduction​ and Contributions

Methodology

​Results and Experiments

​Conclusion​

​

3 of 34

INTRODUCTION & CONTRIBUTIONS

Reconstructing Geometry and BRDF of reflective objects without the need of masks.

​

Calculating both direct and indirect lights for evaluating environment on a reflective surface.

4 of 34

METHODOLOGY

NeRO

4

5 of 34

BREAKDOWN

  1. Preliminaries (NeUS and BRDF)
  2. Geometry Reconstruction
  3. BRDF Estimation

6 of 34

NEUS

 

7 of 34

MICROFACET BRDF

 

8 of 34

STAGE 1: GEOMETRY RECONSTRUCTION

Using NeUS, but with different color shading.

​

Color predictions and assumptions to follow in next slides.

9 of 34

SPLIT-SUM APPROXIMATION

 

 

 

 

 

10 of 34

 

11 of 34

LOSSES (GEOMETRY)

Ray marching along reflection to compute occlusion probability

12 of 34

SIDE NOTE: CAPTURER REFLECTION

 

13 of 34

STAGE 2:�BRDF ESTIMATION

Refining BRDF. We have obtained rough approximation of BRDF until now.

Using Monte Carlo Sampling to obtain colors.

14 of 34

DIFFUSE

 

15 of 34

SPECULAR

 

 

16 of 34

REGULARIZATION

17 of 34

RESULTS AND EXPERIMENTATION

Dataset:

Glossy-Blender Synthetic dataset

Glossy-Real dataset

18 of 34

GLOSSY-BLENDER

19 of 34

GLOSSY-REAL

Painted glossy objects with non-reflective coating to get groundtruth 3D surface

20 of 34

21 of 34

RESULTS ON GLOSSY-BLENDER

22 of 34

Geometry Reconstruction

23 of 34

Relighting

24 of 34

Reconstruction Quality

(Chamfer Distance)

Relighting

(PSNR)

25 of 34

RESULTS ON GLOSSY-REAL

26 of 34

Geometry Reconstruction

27 of 34

Relighting

28 of 34

Reconstruction Quality

(Chamfer Distance)

29 of 34

ABLATION STUDIES

30 of 34

Geometry Reconstruction

31 of 34

32 of 34

DECOMPOSITION

33 of 34

FAILURE CASES

34 of 34

THANK YOU