Leandro

20
A Theory of Shape by Space Carving Kiriakos N. Kutulakos Department of Computer Science and Department of Dermatology, University of Rochester, Rochester, NY Steven M. Seitz The Robotics Institute, Carnegie Mellon University, Pittsburgh Published at the Special Issue on Genomic Signal Processing, 2000.

description

 

Transcript of Leandro

Page 1: Leandro

A Theory of Shape by Space Carving

Kiriakos N. Kutulakos  Department of Computer Science and Department of Dermatology, University of

Rochester, Rochester, NY

Steven M. Seitz  The Robotics Institute, Carnegie Mellon University, Pittsburgh

Published at theSpecial Issue on Genomic Signal Processing, 2000.

Page 2: Leandro

Problem Statement

• Assuming:– Arbitrary (known) camera placement.– Arbitrary scene shape.– Arbitrary scene texture, color and background.– Lambertian.

• Input: N photographs of a scene

• Output: 3D structure of the scene

Page 3: Leandro

Context

• Traditional Approach to 3D Reconstruction– Correspondence based

“Many scenes can account for the input images”

Page 4: Leandro

Shape by Space Carving

• Scene or volume based

• No correspondences needed!

Page 5: Leandro

Picture-Shape Constraint

Page 6: Leandro

Photo Hull

• Definition: Maximal volume, V, reproducing all possible silhouettes of an object, S, from viewing region R.

R

S V

Page 7: Leandro

Photo Hull in 3D

Page 8: Leandro

Shape by Space Carving

• Given a volume containing the object, iteratively carves it out until it converges to the Photo Hull.

Page 9: Leandro

Shape by Space Carving Analogy

Page 10: Leandro

Shape by Space Carving

True Scene Reconstruction

Least-Commitment Recovery

Page 11: Leandro

Carving outHow to decide if a certain voxel belongs to the Photo Hull?

C o n s i s t e

n c y

Page 12: Leandro

ConsistencyRendering colored voxels from input viewpoints reproduces original images.

Point photo-consistent: A point (voxel) that can be assigned a color consistent with its projection in all visible cameras.

shape

Page 13: Leandro

Consistency (cont.)

Shape Photo-consistent: A shape for which all surface points are photo-consistent.

Photo-Hull: Union of all photo-consistent subsets of initial volume.

VPhoto-hull of V

Photo-consistent subsets of V

Page 14: Leandro

Proposed Algorithm

1. Initialize the volume containing the shape;

2. Analyze each visible voxel for point consistency (plane sweep strategy);

3. Repeat step 2 until no voxel is found inconsistent;

4. Merge photo consistent subsets.

Page 15: Leandro

Plane Sweep Strategy

Sequence of plane sweeps along 6 principle directions.

Only cameras behind the sweep plane are used for consistency checking.

Page 16: Leandro

Experimental Results

Page 17: Leandro

Experimental Results

Page 18: Leandro

Experimental Results

Page 19: Leandro

Experimental Results

Page 20: Leandro

Space Carving Discussion

• Least-commitment Reconstruction– No a-priori constraints– Flexible acquisition– Insensitivity to occlusion– Efficient– Scalable

• Setup– Camera calibration– Sensitivity to specularities and transparencies