Supplementary Material of "Look-Ahead Training with Learned Reflectance Loss for Single-Image SVBRDF Estimation"

Xilong Zhou and Nima Khademi Kalantari

In the supplementary material, we compare our results against [Deschaintre18], [Deschaintre19],[Zhou21], [Guo21], [Gao19] and MaterialGAN ([Guo20]) on synthetic data with GT, real data with GT provided by MaterialGAN, real data with GT provided by us and real data without GT



Synthetic data

We show generated feature maps, followed by renderings under a moving light source





Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, both [Guo21] and our approach generate similar renderings, but the normal map of [Guo21] is overly sharp and specular map is too bright compared with GT. Our approach generates better feature maps.


Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, our approach generates feature maps closer to GT compared with all other approaches.


Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours






Real data with GT provided by MaterialGAN

We show generated feature maps, followed by 4 re-rendered images under reference lightings and renderings under a moving light source





Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, only [Guo20] and our method are able to reconstruct highlight region properly in the diffuse map, but our approach produces renderings that match the GT better.


Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours






Real data with GT provided by us

We show generated feature maps, followed by 4 re-rendered images under reference lightings and renderings under a moving light source





Remarks: in the example below, only our approach is able to properly reconstruct the highlights on the golden flowers and produce renderings that match the ground truth.


Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, our approach produces cleaner renderings. Note that only our method is able to properly reproduce the strong highlights on the golden pattern that match the ground truth.


Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, [Zhou21] produces reasonable renderings, but ours better match the ground truth. Note that the their highlights are spread over the entire tile, but our highlights, similar to the ground truth, are more concentrated.


Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, note that the rendered images by [Zhou21] are brighter than the ground truth, and the renderings by [Guo20] have stronger highlights. Ours better match the ground truth.


Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, renderings of both [Zhou21] and ours are visually good, but specular highlight of [Zhou21] is not concentrated.


Input
SVBRDF Renderings
GT
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours






Real data without GT

We show generated feature maps, followed by renderings under a moving light source





Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, [Guo21], similar to our method, captures the difference between the shining metal and the rough stone, but their normal map is overly sharp, introducing unnecessary high-frequency details to the renderings.


Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, note that, [Deschaintre19], [Guo21], and [Gao19] introduce an unnecessary bump in the normal map (in the highlight region). [Guo21] and [Zhou21] produce renderings that are comparable to ours, although our reflectance maps are slightly cleaner.


Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, specular hightlight in red region is overestimated in all other approaches except [Zhou21] and our method, but our method generates cleaner renderings than [Zhou21].


Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, only [Zhou21] and our approach generate reasonable renderings, but the diffuse map of [Zhou21] slightly bakes hightlight in the diffuse map.


Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, in the below example, only our approach properly captures the difference between the roughness of leather and fabric, which can be clearly seen in the renderings.


Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, our approach generates cleaner renderings with moving focused specular highlight, compared with other approaches.


Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours




Remarks: in the example below, most other approaches produce normal maps with unnecessary details. Our approach produces the cleanest renderings and reflectance maps.


Input
SVBRDF Renderings
[Deschaintre18]
[Deschaintre19]
[Zhou21]
[Guo21]
[Gao19]
[Guo20]
Ours