Upload
ansley-channer
View
242
Download
0
Tags:
Embed Size (px)
Citation preview
1
Photometric Stereo Photometric Stereo ReconstructionReconstruction
Dr. Maria E. AngelopoulouDr. Maria E. Angelopoulou
22
Photometric Stereo (PS) Basics
Input images: same viewpoint / different illumination directions
Varying albedo values → minimum of 3 illumination directions
The measured pixel intensities and actual albedo for a given patch o are related as follows
Outputs: orientation & albedo of each surface facet Height map produced after integration of the surface
normals.
(1)
33
Pros & Cons of PS Compared to Conventional Stereo
Pros+ ability to operate on featureless objects+ absence of feature matching errors+ computational simplicity
Cons- controlled imaging conditions- inaccurate simplifying assumptions- need for off-line calibration sessions & exclusive
access to the imaging system
44
Objective
Provide an autonomous, purely data-driven PS reconstruction system that is suitable for real-life applications.
Focus on: uncalibrated flatfielding & uncalibrated light estimation.
Diffuse light component as a degradation factor for intended directional lighting. Indicate up to which ratio of ambient to directional light component photometric stereo gives useful reconstruction outputs.
5
The Inverse-Square Law of Light Propagation
Fundamental assumption of PS: The variation in brightness for a given pixel is solely dependent on the angle between the illumination vector and the surface normal at the corresponding real-world surface facet.
In practice the inverse-square law of light propagation renders the above assumption inaccurate.
To correct the input data appropriately, flatfielding may be employed. Flatfielding employs a set of reference images captured at a dedicated imaging session under the same imaging conditions as the main session. The illumination variations of the reference images is solely due to inhomogeneities of the system.
66
Standard Calibrated Flatfielding Technique
Use a grey piece of card as calibrating device. Photograph the card multiple times under the same
illumination as the main imaging session. Perform 2D 2nd order polynomial fitting on the flatfielding
reference images to smooth out high frequency noise. This gives the illumination fields .
For every point of the image plane, the new pixel value is computed as
(2)
77
Standard Calibrated Flatfielding Technique
88
Uncalibrated Flatfielding
Intensity values decrease across the kth illumination field from the brightest point to the darkest point. Due to the inverse-square law, the 2D illumination field can be approximated with
Only radial distance matters, and thus
Placing the origin of the Cartesian coordinate system at the brightest point:
(3)
(4)
(5)
9
(6)
(7)
Employing constraints (4) and (5), we get:
In (6) the origin of the coordinate system moves to the brightest point of each kth image. Instead of moving the origin for each k, it is more convenient to express (6) on the fixed image coordinate system as:
Uncalibrated Flatfielding
10
Uncalibrated Flatfielding
(8)
(10)(9)
(11)
Find the values of the brightest and darkest points:
Consider the location of the brightest and darkest points:
11
Uncalibrated Flatfielding
1212
Uncalibrated Flatfielding
1313
Uncalibrated Flatfielding
14
PS Reconstruction
(a) no flatfielding, (b) calibrated flatfielding, (c) uncalibrated flatfielding
15
PS Reconstruction
16
- The relationship between surface gradients and brightness is captured by the reflectance map of the surface.
- A Lambertian sphere illuminated by a point source in direction
16
Standard Calibrated Estimation of Illumination Directions
- Let denote the global maximum of the kth
reflectance map. It is
has a reflectance map of the form:
(12)
(13)
17
- If patch o is projected at pixel ,
- Let the gradient at surface patch o be .
17
Standard Calibrated Estimation of Illumination Directions
then the measured image intensity at that pixel is given by the image irradiance equation as
. (14)
- Thus the surface orientation that maximizes the Lambertian reflection component is the one for which the normal vector points to the illumination source!
18
the global maximum of the Lambertian reflectance mapcorresponds to the maximum intensity measurement , theluminous dot. If the latter resides at and is the projection of the real-world patch m then
18
Standard Calibrated Estimation of Illumination Directions
(13) and (14)- Due to equations
- If a Lambertian sphere is photographed, the normal vector canbe recovered at any point and can be estimated once theluminous dot is identified.
(15)
1919
Uncalibrated Estimation of Illumination Directions
The proposed uncalibrated illumination vector technique targets the human face class.
A human face can be realistically approximated with a 3D ovoid that is reconstructed on top of the face area. A 3D ovoid is given in the xyz Cartesian system by
where
2020
Uncalibrated Estimation of Illumination Directions
2121
Uncalibrated Estimation of Illumination Directions
22
PS Reconstruction
23
PS Reconstruction
24
PS Reconstruction
25
The Effect of Diffuse Light on PS Reconstruction
PS requires the subject to be illuminated in turn by directional light sources.
The ambient light component that is present constitutes a degrading factor that reduces the directionality of the intended directional component.
Objectives:– Assess the PS robustness with respect to the ratio of
ambient to directional illuminance. – Find the illuminance ratio where reconstruction is no longer
informative of the actual surface.
26
The Effect of Diffuse Light on PS Reconstruction: Experimental Setup
Floodlight connected to a dimmer provides uniform ambient illumination of varying illuminance.
Light-/Flash-meter used to measure both ambient and directional components.
27
The Effect of Diffuse Light on PS Reconstruction
λ=9
λ=11
28
Conclusions
The proposed uncalibrated flatfielding technique is general-purpose and renders similar reconstruction results as its calibrated counterpart.
The proposed uncalibrated light estimation technique is a practical approach that targets the human face class.
The above techniques enable autonomous and reliable surface reconstruction for challenging real-world applications, such as the on-line capturing of human faces.
In the presence of ambient light, PS reconstruction quality decreases linearly as the illuminance ratio of the diffuse to the directional light component increases. PS provides informative outputs for illuminance ratios as high as λ=9.
2929
Thanks for your attention!