• Aucun résultat trouvé

Local Shape Editing at the

N/A
N/A
Protected

Academic year: 2022

Partager "Local Shape Editing at the"

Copied!
35
0
0

Texte intégral

(1)

Local Shape Editing at the Compositing Stage

Carlos J. Zubiaga, Gaël Guennebaud, Romain Vergne, Pascal Barla

(2)

Compositing

Final Image

Subsurface

Emission

Transparent

Diffuse Reflection

Shading Buffers

(3)

Subsurface

Emission

Transparent

Diffuse Reflection

Shading Buffers

Compositing

Diffuse Coeff.

Ambient Occ.

Auxiliary Buffers

Object ID

Normals Depth/Position

(4)

Compositing Effects

DOF / Fog Relighting

Final Image

Auxiliary Buffers Shading Buffers

Normals Depth/Position

Diffuse Reflection

(5)

Compositing Effects Limitations

Normals

Auxiliary Buffers

Depth/Position

Shading Buffers

Diffuse Reflection

Shape modifications do not change shading Lighting is lost in the rendering process

Costly re-rendering is needed

(6)

Goal

Grant real-time local shape modification in post-processing in a plausible way

(7)

Previous Work

(8)

Lighting Reconstruction

Reflectance and Natural Illumination from a Single Image [Lombardi et al 12]

• Assume natural lighting and low entropy

• Statistical BRDFs and a low detail lighting A Signal-Processing Framework for Inverse Rendering

[Ramammorthi et al 01]

• Inverse Rendering is an ill-Posed Problem

Photograph Render

Env. Light Map Photograph Reconstructed Illumination

Decoupling of lighting and material is

not necessary to manipulate appearance

(9)

Appearance Manipulation I

Image Based Material Editing [Khan et al 06]

• Estimates of 3D shape &

lighting for re-rendering

Interactive Reflection Editing [Ritschel et al 09]

• Manipulate reflections on 3D away from physical restrictions

Input HDR Modification 1 Modification 2

Input Result Result Input

Control Control

Manipulation performed in full 3D:

not adapted to compositing

(10)

Appearance Manipulation II

Surface flows for image-based shading design [Vergne et al 12]

• Use depth and normal buffers to deform/warp images

MatCap Decomposition for Dynamic Appearance Manipulation [Zubiaga et al 15]

• Modify lighting and material appearance from MatCaps

• Avoid decoupling of lighting and material

Texture effects

Add reflections Shinier material

Sahding effects

A novel approach to manipulate existing appearance

in complex renderings is required

(11)

Main Idea

Reconstruct Pre-filtered Environment Maps (PEM) per object/material

(12)

Frequency-dependent Reconstruction Approach

Diffuse

• Low-Frequency

• Spherical Harmonics

Reflections

• High-Frequency

• Detailed PEM

(13)

Reconstruction

(14)

Diffuse Reconstruction

Input

Diffuse Shading

Gauss Map Diffuse PEM Diffuse

reconstructed Residuals Reconstruction

Spherical Harmonics

Fitting

Quadratic Programming (LS + “ ≤ )

Mean Least Square

+1

0

-1

Residuals corresponds to darkening by occlusion

Residuals can be reintroduced

(15)

Reflection Reconstruction

High Resolution Low Resolution

Interpolation

Spherical barycentric coordinates Weighting

Favorize smallest quads

• Blurring

• Discontinuities

(16)

Reflection Reconstruction

Dual Paraboloid Map Front Back

1st Hole Filling

2nd Regularize

Harmonic interpolation on tessellated sphere

Input

Sligthly blur discontinues

of disconnected polygons

(17)

Recontruction Validation

Diffuse

Diff ×5

Render Recontructed Input Ground Truth

Diff ×5 Diff ×100

Diff ×10

Reflections

Render Recontructed Input Ground Truth

(18)

Reconstruction Validation

Normals Shading

Diffuse

Reflection Sphere in

perspective Head

Normals Shading

Vase

Normals

Shading

(19)

Re-compositing

(20)

General Pipeline

Original

PEM Occlusion

Diffuse Reflection

Recomposited

Modified

Recompositing

Combine original & PEM Reconstruction

(21)

Diffuse pipeline

Normal Reconstructed

Occlusion

PEM

Linear interpolation

• Preserved occlusion

Recomposited

Residual

• Re-introduce local darkening

Original

(22)

Reflections pipeline

Reconstructed Occlusion

• Binary Mask

• Disable modification

Recomposited

Diff Normals

Linear interpolation

• Avoid ghosting effects

Orig. Mod. PEM

Original

(23)

Recap

Interpolate using occclusion

Interpolation

Diffuse

Add residuals

Reflections

(24)

Results

(25)
(26)

Normal Modification Comparison

Ours Ground Truth Ours Ground Truth

(27)
(28)
(29)
(30)

Complex Scene Reconstruction

(31)
(32)

Timings (ms)

Model Diffuse

Reconstruction

Reflection Reconstruction

Partial Reconstruction Hole filling

Red Sphere 110 315 215

Red Head 184 470 280

Red Vase 147 385 192

Cup 29 45 325

Kettle 65 97 650

Black Vase 64 133 4000

Truck Body 119 120 6000

Front Mudguard 25 45 2500

Rear Mudguard 14 26 3700

(33)

Conclusions

Restricted to opaque objects

Limited by the geometric complexity Normal mapping, not displacement

Plausible pre-filtered env. maps are sufficient to modify local shape at the compositing stage

Limitations

Extension to spatially-varying reflectance Output more info at the rendering stage Combine with manipulation of materials

Future Work

(34)

Thank you

http://prism-network.eu/

http://gratin.gforge.inria.fr/

https://www.thefoundry.co.uk/

(35)

Complex Scenes Input

Normal Specular

Shading

Surface ID Ambient

Occlusion Specular

Occlusion

Diffuse

Shading

Références

Documents relatifs

Etant conscients de ces difficultés, nous supposerons donc dans cet article que tous les paramètres de la source de lu- mière, de la réflectance de la surface et de l’appareil

SFS method is extremely robust to pixel noise: figure 3-b) displays the re- sult produced by this algorithm (after 10 iterations) using the image of a text page with its

Kyu Park, “Robust Light Field Depth Estimation for Noisy Scene With Occlusion,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp..

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des

Finally, this chapter considers the different strategies of teaching, learning and testing vocabulary, discusses its role and integration in an ESP course and finally draws

By analyzing how different frequency com- ponents of the recordings reach the various microphones through time, it extracts both spatial information and audio content for the

The advected smoke layers over Europe are detected at both tropospheric and stratospheric heights, with the latter presenting non-typical values of the Linear

l’approche que nous avons adoptée, par deux voies diffé- rentes, consiste à supposer que la surface est localement sphérique : dans le paragraphe 3, cette hypothèse est ex- plicite