Computer Vision: Shape from Specularities and MotionPresentation Transcript
Use of Specularities and Motion in the Extraction of Surface Shape Damian Gordon [email_address]
Introduction
Introduction - Image Geometry
Photometric Stereo (1)
Structured Highlights (1)
Stereo Techniques (2)
Motion Techniques (3)
Solder Joint Inspection (1)
Specular Surface
Angle of Incidence = Angle of Reflection
Image Geometry ________________________
Image Formation
Geometry - determines where in the image plane the projection of a point in a scene will be located
Physics of Light - determines the brightness of a point in the image plane as a function of scene illumination and surface properties
Image Formation
Image Formation
The LINE OF SIGHT of a point in the scene is the line that passes through the point of interest and the centre of projection
The above model leads to image inversion, to avoid this, assume the image plane is in front of the centre of projection
Image Formation
Perspective Projection
(x’,y’) may be found by computing the co-ordinates of the intersection of the line of sight, passing thru’ (x,y,z) and the image plane
By two sets of similar triangles :
x’=fx/z and y’=fy/z
Image Irradiance (Brightness)
The irradiance of a point in the image plane E(x’,y’) is determined by the amt of energy radiated by the corresponding scene in the direction of the image point : E(x’,y’)=L(x,y,z)
Two factors determine radiance emitted by a surface patch
I) Illumination falling on scene patch
- determined by the patch’s position relative to the distribution of light sources
II) Fraction of incident illumination reflected by patch
- determined by optical properties of the patch
Image Irradiance
( i i is the direction of the point source of scene illumination
( e e is the direction of the energy emitted from the surface patch
E( i i is the energy arriving at a patch
L( e e is the energy radiated from the patch
Image Irradiance
The relationship between radiance and irradiance may be defined as follows :
L( e e f( i i e e E( i i
where f( i i e e is the bidirectional reflectance distribution function (BRDF)
BRDF - depends on optical properties of the surface
Types of Reflectance
Lambertian Reflectance
Specular Reflectance
Hybrid Reflectance
Electron Microscopy Reflectance (not covered)
Lambertian Reflectance
Appears equally bright from all viewing directions for a fixed illumination distribution
Does not absorb any incident illumination
BRDF is a constant (1/ )
Lambertian Reflectance - Point Source
Perceived brightness illuminated by a distant point source
L( e e Cos s -- Lambert Cosine Rule
this means, a surface patch captures the most illumination if it is orientated so that the surface normal of the patch points in the direction of illumination
Lambertian Reflectance - Uniform Source
Perceived brightness illuminated by a uniform source
L( e e
this means, no matter how a surface is illuminated, it receives the same amount of illumination
Specular Reflectance
Reflects all incident illumination in a direction that has the same angle with respect to the surface normal, but on the oppside of the surface normal
light in the direction ( i i is reflected to ( e e ( i i +
BRDF is ( e - i ( e - i - / Sin i Cos i
Specular Reflectance
Perceived brightness is
L( e e e e
this means, the incoming rays of light are reflected from the surface like a perfect mirrior
Hybrid Reflectance
Mixture of Lambertian and Specular reflectance
BRDF is
( e - i ( e - i - / Sin i Cos i
where is the mixture of the two reflectance functions
Surface Orientation
If (x,y,z) is a point on a surface and (x,y) is the same point on the image plane, with distance z from the camera (depth), then a nearby point is
(x+ x, y+ y)
the change in depth can be expressed as
z = ( z/ x) x + ( z/ y) y
Surface Orientation
The size of the partial derivaties of z with respect to x and y are related to the orientation of the surface patch.
The gradient of (x,y,z) is the vector (p,q) which is given by
p = ( z/ x) q ( z/ y)
Reflectance Map
For a given light source distribution and a given surface material, the reflectance of all surface orientations of p and q can be catalogued or computed to yield the reflectance map R(p,q) which leads to the image irradiance equation
E(x,y) = R(p,q)
Reflectance Map
i.e., that the irradiance at a point in the image plane is equal to the reflectance map value for surface orientation p and q in the corresponding point in a scene
in other words, given a change in surface orientation, the reflectance map allows you to calculate a change in image intensity.
Shape from Shading
the oppside problem, we know E(x,y) = R(p,q), so we need to calculate p and q for each point (x,y) in the image
Two unknows, one equation, therefore, a constraint must be applied.
Shape from Shading
Smoothness constraint
Objects are made of a smooth surface, which depart from smoothness only along their edges
may be expressed as
Shape from Shading
Photometric Stereo
Asssume a scene with Lambertian reflectance
Each point (x,y) will have brightness E(x,y) and possible orientations p and q for a given light source
if the same surface is illuminated by a point source in a different location, the reflectance map will the different
Photometric Stereo
Using this method, surface orientation may be uniquely identified
In reality, not all incident light is radiated from a surface, this is accounted for by adding an albedo factor ( ) into the image irradiance eqn.
E(x,y) = R(p,q)
Photometric Stereo ________________________
Determining Surface Orientations of Specular Surfaces by Using the Photometric Stereo Method Katsusi Ikeuchi Ministry of International Trade in Industry, Japan
Introduction
Photometric stereo may be used to determine the surface orientation of a patch
for diffuse surfaces, point source illumination is used
for specular surfaces, a distributed light source is required
Image Radiance
For a specular surface and an extended light source :
L e ( e e L i e e
Relationship between reflected radiance and image irradiance
E p = {( /4)(d/f p ) 2 Cos 4 }L e
fp = focal length
d = diameter of aperture
= off-axis angle
Image Radiance
from this a brightness distribution may be derived
and from that an inverse transformation
System Implementation
Two Stage Process
Off-Line Job
On-Line Job
Off-Line Job
Light Source : Three linear lamps, placed symmetrically 120 degrees apart
Lookup Table : Could use 3D table, but observed triples often contain errors
Instead use 2D lookup Table - each element has two alternatives
Each alternative consists of a surface orientation and an instensity
Off-Line Job
On-Line Job
Normalization is required to cancel the effect of albedo
Brightness calibration is required also
The correct alternative of the two solutions is found by comparing the distance between the actual third image brightness and the element of the matrix
Results
Works well in a contrainted environment
has problems if the surface is not smooth
Extracting the Shape and Roughness of Specular Lobe Objects Using Four Light Photometric Stereo Fredric Solomon Katsushi Ikeuchi Carnegie Mellon
Structured Highlights __________________________
Structured Highlight Inspection of Specular Surfaces Arthur C. Sanderson Lee E. Weiss Shree K. Nayar Carnegie Mellon
Introduction
Structured Highlight approach yields 3D images from point sources and images
‘ Highlight’ - light source reflected on a specular surface
Introduction
Angle of Incidence = Angle of Reflection
A fixed camera will image a reflected light ray (highlight) only if it is positioned and orientationed correctly
Introduction
Once a highlight is observed, if the direction of the incident ray is known, the orientation of the surface element may be found
A spherical array of fixed point light sources is used to ensure all positions and directions are scanned
Lambertian Reflectance
The reflectance relationship for a Lambertian model of image E(x,y)
E(x,y) = A (n . s)
n = surface normal (unit vector)
s = source direction (unit vector)
A = constant related to illumination intensity and surface albedo
Hybrid Reflectance
The reflectance relationship for a hybrid model of image E(x,y)
E(x,y) = A k (n . s) + (a/2)(1-k) .
[2(n . z)(n . s)-(z . s)]
z = viewing direction (unit vector)
k = relative weight of specular and Lambertian components
n = sharpness of the specularity
Structured Hightlight Inspection
Using the above equation, the slope of any point may be calculated
Surface orientation may be determined by the sources that produce local peaks in the reflectance map.
Camera Models
Perspective Camera Model
Orthographic Projection Model
“ Fixed” Camera Model
Perspective Camera Model
All reflected rays pass though a focal point
this model provides very accurate measurements, but requires extensive calibration procedures
Orthographic Projection Model
the focal point is assumed to be an infinite distance from the camera and all the reflected rays are perpendicular to the image plane
“ Fixed” Camera Model
all rays are emitted from a single point on the reflectance plane and all surface normal estimates are computed to that reference point
Camera Models - Accuracy
Perspective Camera Model
Most accurate
“ Fixed” Camera Model
Next most accurate
Orthographic Projection Model
Most sensitive to error
SHINY - Structured Highlight INspection sYstem
Highlightrs are extracted from images and tablulated
Surface normals are computed based on lookup tables dervied from calibration experiments
Reconstruction is done using interpolation followed by smoothing
Stereo Hightlight Algorithm
The assumption of a distant source to uniquely identify the angle of incidence of illumination is an approximation
To improve this, a second camera is used with stereo matching for greater accuracy
Results
With two cameras need to resolve stereo matching ambiguities, therefore, need further constraints
This technique is slow (1988)
Stereo Techniques ________________________
Stereo in the Presence of Specular Reflection Dinkar N. Bhat Shree K. Nayar Columbia University
Introduction
Stereo is a direct method of obtaining the 3D structure of the visual world
But, it suffers from the fact that the correspondence problem is inherently underconstrained
Correspondence Problem
the most common constraint is that intensities of corresponding points in images are identical
The assumption is not valid for specular surfaces (since intensity is dependant on viewing direction)
Specular Reflection
When a specular surface is smooth, the distribution of the specular intensity is concentrated
As the surface becomes rougher, the peak volume of the specular intensity decreases and the distribution widens
Specular Reflection Smooth Surface Rough Surface
Implications for Stereo
The total image intensity of any point is the sum of the diffuse and specular intensity conponents
Since the change in diffuse components is very small relative to the changes in specular components, it follows that the overall change in intensity is approximately equal to the specular intensity differences
I diff ~= | I s1 - I s2 |
Implications for Stereo
This approximation will assist in determining an optimal binocular stereo configuration, which minimises specular correspondence problems but maximises precision in depth estimation
Binocular Stereo Configuration
Vergence
When cameras are orientated such that their optical axes intersect at a point in space, this point is refered to as the point vergence
Depth accuracy is directly proportional to vergence (…which conflicts with the requirement to minimize intensity differences)
Binocular Stereo
Determining the maximum acceptable vergence can be formulated as a constrained optimization problem
f obj = v 1 . v 2
c 1 : I diff < a specified threshold
c 2 : the cameras lie in the X-Z plane
Experiments
Two uniformly rough cylindrical objects wrapped, one is gift wrapper and the other in xerox paper
Similar patterns were marking on both
Trinocular Stereo
Required in environments which are less structured and where surface roughness cannot be estimated
Allows intensity difference at a point to be constrained to a threshold in at least one of the stereo pairs
Trinocular Stereo
Experiments
The experiments done indicate that the reconstruction algorithm works resonably well in an unconstrained environment
Retrieving Shape Information from Multiple Images of a Specular Surface Howard Schultz University of Massachusetts
Introduction
This research extends a diffuse mutli-image shape-from-shading technique to perform in the specular domain
Viewing Geometry
Assumes an ideal camera with focal length f viewing a surface
The camera focal point is located at P and O is a point on the surface
From Snell’s Law an equation can be derived relating the objects position in space to its image on the image plane
Viewing Geometry
Image Synthesis
the specular surface stereo method requires a model that predicts accurately the irradiance at each pixel
Use Idealized Image Synthesis Model
this will allow us to determine that the irradiance is directly proportional to the product of the radiance and the reflection co-efficient
Specular Surface Stereo
Starting at a known evelation, an iterative process is used to determine shape
Two-step process, determine orientation and propagation
Surface Orientation
Identify the pixels that view the surface point (by calculating an inverse of a projective transform)
A value of (p,q) is found such that the predicted irradiance at E(p,q) match the observed values
Surface Propagation
if a point is known on a surface, it is possible to recover shape by propagation
If (x,y) has elevation h and gradient (p,q) then (x+ x, y+ y) has elevation
h’ = h +p x +q y
Obtaining Seed Values
if there are surface features with diffuse proprties (e.g. scratchs or rough spots), use feature matching methods
if surface is smooth, use a laser range finder
Results
Tests were done on four _simulated_ images to determine the feasibilty of the method, the results were 99% accurate
Using this method in the ‘real world’ would require more constraints
Motion Techniques ________________________
A Theory of Specular Surface Geometry Michael Oren Shree K. Nayar Columbia University
Introduction
Develops a 2D profile recovery technique and generalize to 3D surface recovery
Two major issues associated with
specular surfaces
detection
shape recovery
Introduction
Specular surfaces introduce a new kind of image feature, a virtual feature
A virtual feature is the reflection by a specular surface of another scene point which travels over the surface when the observer moves.
Curve Representation
Cartesian co-ordinates result in complex equations describing specular motion
Using the Legendre transform to represent the curve as an envelope of tangents
Curve Representation
2D Caustics
When a camera moves around an object the virtual features move on the specular surface, producing a family of reflected rays (the envelope defined by this family is called the caustic )
On the other hand, the caustic of a real feature is one single point (the actual position of the feature in the scene where all the reflected rays intersect)
Test Image
2D Caustics
Using this, feature classification is simply a matter of computing a caustic and determining whether it is a point or a curve
Features are tracked from one frame to the next using a sum of square difference (SSD) correlation operator
2D Profile Recovery
The camera is moved in the plane of the profile and the features are tracked
An equation may be derived relating the caustic to the surface profile, allowing the recovery of the 2D profile from the image.
3D Surface Recovery
The 3D camera motion problem will result in an arbitrary space curve rather than a family of curves as in the 2D case
The 3D problem cannot but reduced to a finite number of 2D profile problems
3D Surface Recovery
The concept behind the derevation of the 3D caustic curve is to decompose the caustic point position at any given instant into two orthogonal components
As the camera moves along the specular object, a virtual feature travels along the 3D profile on the objects surface.
It is possible to develop an equation which relates the trajectory of the virtual feature to the surface profile
Results
The 2D testing involved tracking two features on two different specular surfaces, in both experiments the profile was accurately estimated
The 3D testing involved tracking a highlight on a specular surface, the recovered curve is in strong agreement with the actual surface
Epipolar Geometry ________________________
Epipolar Geometry
two cameras are displaced from each other by a baseline distance
Object point X forms two distinct image points x and x’
Epipolar Geometry
Assume images formed in front of camera to avoid inversion problem
point (x’, y’) in the images plane from a real point (x, y, z) may be calculated as
x’ = fx/z and y’ = fy/z
the displacement between the locations of image point is called the disparity
Epipolar Geometry
the plane passing through the two camera centres and the object point is called the epipolar plane
the intersection of the image plane and the epipolar plane is called the epipolar line
Generalizing Epipolar-Plane Image Analysis on the Spatiotemporal Surface H. Harlyn Baker Robert C. Bolles SRI International
Introduction
The technique of Epipolar-Plane Image Analysis involves obtaining depth estimates for a point by taking a large number of images
This gives a large baseline and higher accuracy
It also minimises the correspondence problem
Epipolar-Plane Image Analysis
this technique imposes the following constraints
the camera is moving along a linear path
it acquires images at equal spacing as it is moved
the camera’s view is orthogonal to the direction of travel
Epipolar-Plane Image Analysis
the traditional notion of epipolar lines is generalized to an epipolar plane
using this, plus the fact that the camera is always moving along a linear path and we may conclude the a given scence feature will always be restricted to a given epipolar plane
Epipolar-Plane Image Analysis
The Spatiotemporal Surface
As images are collected, they are stacked up into a spatiotemporal surface
as each new image is obtained its spatial and temporal edge contours sre constructed
using a 3D Laplacian of a 3D Gaussian
The Spatiotemporal Surface
3D Surface Estimation and Model Construction From Specular Motion in Image Sequences Jiang Yu Zheng Norihiro Abe Kyushu Institiute of Technology Yoshihiro Fukagawa Torey Corporation
Introduction
This technique reconstructs 3D models of complex objects with specular surfaces
The process involves rotating the object under inspection
System Setup
Projected Highlights
An extended light source project highlight stripes onto the object
The stripes gradually shift across the object surface and pass most point once
The specular motion is captured in epipolar-plane images
Feature tracking
We know how to detect corners and edge of surface patterns
The motion type of highlights in EPI can be used to determine five categories of shape
convex corner
convex
planer
concave
concave corner
EPI-Plane Images
During the rotation, highlights will split and merge, appear and dissapear, etc.
Results
Using EPIs results in very accurate reconstruction of surface shapes
Visual Inspection System for the Classification of Solder Joints Tae-Hyeon Kim Young Shik Moon Sung Han Park Hanyang University Kwang-Jin Yoon LG Industrial Systems
Introduction
Uses three layers of ring shaped LED arrays, with different illumination angles
Solder Joint are segemented and classified using either their 2D features or their 3D features
Classification of Joints
Preprocessing
Objective is to identify and segement the soldered regions
Solder is isolated both vertically and horozontally
Feature Extraction - 2D
Average gray level value of I 1 and I 3
X 1 = 1/N * I K (x,y)
Percentage of highlights of I 1 and I 2
X 2 = 1/N * U(x,y) * 100
U(x,y) = thresholded image of I 1
Feature Extraction - 3D
Shape recovery is done using a hybrid reflectance model for all samples not in the confidence interval
A reflectance map is built up representing intensity values as a function of orientation for each illumination angle
For each point, three intensity values are recovered and from these and the reflectance map, the orientation is estimated