Integrating Shape-from-Shading & Stereopsis

Research output: ThesisDoctoral Thesis

Abstract

This thesis is concerned with inferring scene shape by combining two specific
techniques: shape-from-shading and stereopsis. Shape-from-shading calculates shape using the lighting equation, which takes surface orientation and lighting information to irradiance. As irradiance and lighting information are provided this is the problem of inverting a many to one function to get surface orientation. Surface orientation may be integrated to get depth. Stereopsis
matches pixels between two images taken from different locations of the same scene - this is the correspondence problem. Depth can then be calculated using camera calibration information, via triangulation. These methods both fail for certain inputs; the advantage of combining them is that where one fails the other may continue to work. Notably, shape-from-shading requires a smoothly shaded surface, without texture, whilst stereopsis requires texture - each works where the other does not. The first work of this thesis tackles the problem directly. A novel modular solution is proposed to combine both methods; combining is itself done using Gaussian belief propagation. This modular approach highlights missing and weak modules; the rest of the thesis is then concerned with providing a new module and an improved module. The improved module is given in the second research chapter and consists of a new shape-from-shading algorithm. It again uses belief propagation, but this time with directional statistics to represent surface orientation. Message passing is performed using a novel method; it is analytical, which makes this algorithm particularly fast. In the final research chapter a new module is provided, to estimate the light source direction. Without such a module
the user of the system has to provide it; this is tedious and error prone, and
impedes automation. It is a probabilistic method that uniquely estimates the light source direction using a stereo pair as input.
LanguageEnglish
QualificationPh.D.
Awarding Institution
  • University of York
Supervisors/Advisors
  • Wilson, Richard C, Supervisor, External person
Award date1 Jan 2009
StatusPublished - 2009

Fingerprint

Lighting
Light sources
Textures
Message passing
Triangulation
Automation
Pixels
Cameras
Statistics
Calibration

Cite this

Integrating Shape-from-Shading & Stereopsis. / Fincham Haines, Tom.

2009. 247 p.

Research output: ThesisDoctoral Thesis

@phdthesis{53ce5001af54496a9b51562da9d09c38,
title = "Integrating Shape-from-Shading & Stereopsis",
abstract = "This thesis is concerned with inferring scene shape by combining two specifictechniques: shape-from-shading and stereopsis. Shape-from-shading calculates shape using the lighting equation, which takes surface orientation and lighting information to irradiance. As irradiance and lighting information are provided this is the problem of inverting a many to one function to get surface orientation. Surface orientation may be integrated to get depth. Stereopsismatches pixels between two images taken from different locations of the same scene - this is the correspondence problem. Depth can then be calculated using camera calibration information, via triangulation. These methods both fail for certain inputs; the advantage of combining them is that where one fails the other may continue to work. Notably, shape-from-shading requires a smoothly shaded surface, without texture, whilst stereopsis requires texture - each works where the other does not. The first work of this thesis tackles the problem directly. A novel modular solution is proposed to combine both methods; combining is itself done using Gaussian belief propagation. This modular approach highlights missing and weak modules; the rest of the thesis is then concerned with providing a new module and an improved module. The improved module is given in the second research chapter and consists of a new shape-from-shading algorithm. It again uses belief propagation, but this time with directional statistics to represent surface orientation. Message passing is performed using a novel method; it is analytical, which makes this algorithm particularly fast. In the final research chapter a new module is provided, to estimate the light source direction. Without such a modulethe user of the system has to provide it; this is tedious and error prone, andimpedes automation. It is a probabilistic method that uniquely estimates the light source direction using a stereo pair as input.",
author = "{Fincham Haines}, Tom",
year = "2009",
language = "English",
school = "University of York",

}

TY - THES

T1 - Integrating Shape-from-Shading & Stereopsis

AU - Fincham Haines,Tom

PY - 2009

Y1 - 2009

N2 - This thesis is concerned with inferring scene shape by combining two specifictechniques: shape-from-shading and stereopsis. Shape-from-shading calculates shape using the lighting equation, which takes surface orientation and lighting information to irradiance. As irradiance and lighting information are provided this is the problem of inverting a many to one function to get surface orientation. Surface orientation may be integrated to get depth. Stereopsismatches pixels between two images taken from different locations of the same scene - this is the correspondence problem. Depth can then be calculated using camera calibration information, via triangulation. These methods both fail for certain inputs; the advantage of combining them is that where one fails the other may continue to work. Notably, shape-from-shading requires a smoothly shaded surface, without texture, whilst stereopsis requires texture - each works where the other does not. The first work of this thesis tackles the problem directly. A novel modular solution is proposed to combine both methods; combining is itself done using Gaussian belief propagation. This modular approach highlights missing and weak modules; the rest of the thesis is then concerned with providing a new module and an improved module. The improved module is given in the second research chapter and consists of a new shape-from-shading algorithm. It again uses belief propagation, but this time with directional statistics to represent surface orientation. Message passing is performed using a novel method; it is analytical, which makes this algorithm particularly fast. In the final research chapter a new module is provided, to estimate the light source direction. Without such a modulethe user of the system has to provide it; this is tedious and error prone, andimpedes automation. It is a probabilistic method that uniquely estimates the light source direction using a stereo pair as input.

AB - This thesis is concerned with inferring scene shape by combining two specifictechniques: shape-from-shading and stereopsis. Shape-from-shading calculates shape using the lighting equation, which takes surface orientation and lighting information to irradiance. As irradiance and lighting information are provided this is the problem of inverting a many to one function to get surface orientation. Surface orientation may be integrated to get depth. Stereopsismatches pixels between two images taken from different locations of the same scene - this is the correspondence problem. Depth can then be calculated using camera calibration information, via triangulation. These methods both fail for certain inputs; the advantage of combining them is that where one fails the other may continue to work. Notably, shape-from-shading requires a smoothly shaded surface, without texture, whilst stereopsis requires texture - each works where the other does not. The first work of this thesis tackles the problem directly. A novel modular solution is proposed to combine both methods; combining is itself done using Gaussian belief propagation. This modular approach highlights missing and weak modules; the rest of the thesis is then concerned with providing a new module and an improved module. The improved module is given in the second research chapter and consists of a new shape-from-shading algorithm. It again uses belief propagation, but this time with directional statistics to represent surface orientation. Message passing is performed using a novel method; it is analytical, which makes this algorithm particularly fast. In the final research chapter a new module is provided, to estimate the light source direction. Without such a modulethe user of the system has to provide it; this is tedious and error prone, andimpedes automation. It is a probabilistic method that uniquely estimates the light source direction using a stereo pair as input.

M3 - Doctoral Thesis

ER -