1 Novel view synthesis from still pictures by Frédéric ABAD under the supervision of Olivier FAUGERAS 1 Luc ROBERT 2 Imad ZOGHLAMI 2 1 ROBOTVIS Team INRIA.

1

Novel view synthesis from still pictures

by

Frédéric ABAD

under the supervision of

Olivier FAUGERAS1 Luc ROBERT2 Imad ZOGHLAMI2

1 ROBOTVIS Team INRIA Sophia Antipolis 2 REALVIZ SA

2

Novel view synthesis

Given data: Few reference photographs Reference camera calibration

Objective: Photo-realistic image generation Free virtual camera motion In particular: correct handling of

parallax and image resolution

3

Novel view synthesis

Usual approaches: Model-based rendering

(light simulation with mathematical models)

Image-based rendering(image interpolation)

Our approach: Hybrid image-model based rendering

(texture mapping)

4

Our approach

Based on a hybrid scene representation• Rough 3D model + few images (reference images and

masks)• Layer factorization

Rendering engine (main processing step)• View-dependent texture mapping• Double layered-structure

Refinement step (post-processing step)• Rendering errors occur when the 3D model is too rough

Mask extraction (pre-processing step)• Segmentation of the layers in the reference images

5

Our approach

Hybrid scene representation

Rendering engine (main processing step)

Refinement step (post-processing step)

Mask extraction (pre-processing step)

6

Our approach





7

Scene representation

Hybrid representation:

Few reference images

Rough 3D model (built by image-based modeling)

3D structure decomposed into layers

Binary layer masks extracted from the reference images

8

Scene representation

9

Scene representation (example)

Reference images

10


3D model

11


Layer map

12


Masks extracted from reference image #1

13

Our approach





14

Our approach





15

Rendering engine

View-dependent texture mapping [Debevec:96]

• Efficient combination of the different reference images with respect to the virtual viewpoint.

• Optimal image resolution

Double layered-structure, three steps:• Independant rendering of each geometric layer with the

best 3 reference textures• Intra-layer compositing (for VDTM)• Inter-layer compositing (for occlusion processing)

16

View-dependent texture mapping

Basic texture mapping Reference image weighting

17

Double layered-structure

Inter-layer compositing

Intra-layercompositing

Intra-layercompositing

18

Rendering engine (example)

Results: hole-filling by VDTM

19

Rendering engine (example)

Results: generated movie

20

Our approach





21

Our approach





22

Refinement stepRendering errors occur with basic texture

mapping if the 3D model is too rough(‘Geometric Rendering Errors’ or GRE)

GRE’s are responsible for ‘ghosting artefacts’ with view-dependent texture mapping

23

Refinement step

Origin of the Geometric Rendering Errors

24

Refinement step

Origin of the ghosting artefacts

25

Refinement stepOur correcting approach:1) Detect GRE’s in auxiliary reference images2) Propagate them in new generated images3) Correct them by image morphing

26

Our correcting approach

Step 1: detect GRE’s in an auxiliary image Model-based stereo [Debevec:96]

27


Step 2: GRE propagation by point prediction

Knownpoint

Knownpoint

Knownpoint

Knownpoint

Searchedpoint

28

Point prediction methods

Example 1: epipolar transfer

29


Example 2: Shashua’s cross-ratio method

[Shashua:93]

30


Other point prediction methods:

3D point reconstruction and projection Trifocal transfer Compact cross-ratio method Irani’s parallax-based multi-frame

rigidity constraint method

31


Step 3: correct the GRE’s by image morphing

32


Experimental comparison of point prediction methods: Epipolar transfer: simplest implementation

but imprecise and instable close to the trifocal plane

Irani’s approach: complex, imprecise and instable method

Cross-ratio approaches: simple, precise and stable methods

33


Experimental application to deghosting

34


Before deghosting

+

35


After deghosting

+

36


Comparison before/after

Before deghosting After deghosting

37

Our approach





38

Our approach





39

Mask extraction

Extract layer masks from reference images

Reference image Ii Layer Cj Mask Mij

40

Mask extraction

Region-based image segmentation: pixel labelling by energy

minimization

Energies

Optimization techniques

41

Mask extraction


minimization

Energies


42

Mask extraction

Energies:

Data attachment term+

Regularization term

43

Energies

Data attachment term

Ensures the adaptation of the labelling to the observed data in the image

Inverse of the labelling likelihood

44

Data attachment term

Usual segmentation criteria: Luminance: luma-key Color: chroma-key Texture: texture-key

Emphasis on a new geometric

criterion: planar-key

45

Regularization termEnsures one stable and unique

solution‘Markovian Random Field’ a priori

4-connexity neighborhoodSecond order cliquesGeneralized Potts Model potential function

46

Exploits geometric a priori knowledge:Scene made with planar patches

(3D model = triangular mesh)

1 label = 1 plane = 1 homography(between the image to segment and an auxiliary image)

Data attachment energy: dissimilarity between the labelled pixel and its

image by the homography associated with the label

Planar-key

47

Planar-key

Dissimilarity(p,HC.p) < Dissimilarity(p,HA.p)

Dissimilarity(p,HC.p) < Dissimilarity(p,HB.p)

D(C,p) < D(A,p)

D(C,p) < D(B,p)

48

Planar-keyExample:

Auxiliary image Main image

Segmented imageStructure of the scene

49

Planar-key

Technique more complex than it seems: Dissimilarity measures Photometric discrepancy robustness Geometric inaccuracy robustness Occlusion shadow error management

50

Dissimilarity measures

51

Dissimilarity measures

*** vuLRGBY YZNSSD ..55..

Main imageAuxiliary image

52

Photometric robustness

Problems when the main and auxiliary Camera Response Functions do not match.

Proposed solution: Piece-wise luminance histogram correction Affine transformation included in the

dissimilarity computation

53

Photometric robustness

Main image Auxiliary image Reference auxiliary image

Initial segmentation Corrected segmentation Reference segmentation

54

Geometric robustness

Problems when homographies are not accurate

Proposed solution: best-match point selection

55

Geometric robustness

Main image Auxiliary image

Reference segmentationWithout processing With processing

56

Occlusion shadow errors

Origin of the occlusion shadow errors:

pixels corresponding to occluded 3D points in auxiliary images

planar-key no longer valid potential labelling errors

(foreground label instead of background label)

57


Example of correction procedure

58


Auxiliary image Main image

With labelling errors Labelling errors corrected

59

Mask extraction


minimization

Energies


60

Mask extraction


minimization

Energies


61


Usual approaches Deterministic (snakes) Stochastic (simulated annealing)

New approach: Graph-based techniques: graph-cuts

[Boykov-etal:99]

62


Graph-cuts: Well-adapted to our problem (MAP computation) Efficient (global minimum for 2-label problem) Low complexity algorithms Many implementations but no practical study

Complete study of the graph-cut implementations

63

Graph-cuts

Minimal energy = minimal cut = optimal labelling

Minimal cut = maximal flow (maxflow-mincut theorem [Ford-Fulkerson:56])

64

Graph-cuts

Generic max-flow algorithms:

65

Graph-cuts

Best implementations:

Shortest augmenting path with

Optimized stopping conditionGeometric optimization

FIFO preflow-push with

Optimized FIFO emptying condition(no efficient geometric optimization)

66

Graph-cuts

Practical implementation speeding-up

67

Conclusion

Contributions: Complete processing chain:

Pre-processing step: new mask extraction technique

(planar-key + graph-cut techniques)

Main processing step: efficient and open rendering framework

(layered structure + view-dependent texture mapping)

Post-processing step: original refinement approach(Geometric Rendering Errors propagation + deghosting)

68

Conclusion

Future work: Mask extraction:

• Other regularization a priori (Chien model)• Sub-pixel segmentation (partial transparency)

Rendering step:• Rendering speed (interactive frame-rate)• Automatization (weighting and ordering schemes)

Refinement step:• ‘3D Image Warping’-like refinement equation • Automatization and integration in the rendering

engine

1 Novel view synthesis from still pictures by Frédéric ABAD under the supervision of Olivier FAUGERAS 1 Luc ROBERT 2 Imad ZOGHLAMI 2 1 ROBOTVIS Team INRIA.

Documents

scene representation

occlusion processing

vdtm slide

image resolution slide

realviz sa slide

generated movie slide

images reference images

different reference