finding glass
DESCRIPTION
Kenton McHenry Jean Ponce David Forsyth. Finding Glass. Background. Layer Seperation (Szleski, Avidan, and Aniandan, CVPR'00), (Levin, Zomet, and Weiss, CVPR'04). - PowerPoint PPT PresentationTRANSCRIPT
Finding Glass
Kenton McHenry
Jean Ponce
David Forsyth
Background
Layer Seperation (Szleski, Avidan, and Aniandan, CVPR'00),
(Levin, Zomet, and Weiss, CVPR'04)
3D Structure (Hata, Saitoh, Kumamura and Kaida, ICPR'96)
(Ben-Ezra and Nayar, ICCV'03)
(Miyazaki, Kagesawa and Ikeuchi, ICCV'03)
(Murase, ICCV'90)
Recognition (Osadchy, Jacobs, and Ramamoorthi, ICCV'03)
Segmentation (Singh and Huang, CVPR'03)
(Adelson and Anandan, AAAI'90)
I = IB+ e
0 < ≤ 1e ≥ 0
Classifying Junctions
Non-Reversing: transparency, ambiguous depth ordering
Double-Reversing: no transparency
Single-Reversing: transparency
(Singh and Huang, CVPR'03)
(Singh and Huang, CVPR'03)
Our Goal
The Background
The appearance of a glass object changes with
the background (i.e. the scene w/o any
transparent objects)We have seen how knowledge of the
background can be extremeley useful in
reconstructing transparent surfacesIdeal situation: know the background, use
background subtraction
Glass Objects and their Edges
Why?HighlightsMirrorsHysteresis
Adelson et al Revisited
Though they focus on junctions they are
classifying edgesThe proposed rules are binary cues between a
transparent object and its background
Proposed Method
Break edges into small segments and classify them
based on the information from the two sidesProperties of glass: transparency, refraction and
reflection
Cues
Transparency Color Similarity Overlay Consistency
Refraction Texture Distortion Blurring
Reflection Highlights
Color Similarity
(HSV) Hue(HSV) Saturation
Overlay Consistency
Texture Distortion
Filer Bank: 2
scales, 6
orientations (0,)
Blurring
DCTShift in mean in
frequency space
Highlights
Highlights on smooth
shiny surfaces tend to
have a profile with a
sharp spike (Healey and Binford, '87),
(Nayar, Ikeuchi and Kanade, '91)
Highlights
Iteratively fit a line to
perimeter (starting
from threshold of 1.0)Plot line fit errors
Highlights
Single Classifier
5 cues provide 6 valuesSVM with Gaussian kernelMust be conservative with false positives
Classifier can achieve high accuracy on training
data Move hyperplane until true positives < 30%
Multiple Classifiers
glass ⇐ similar_color ∧ high_alpha ∧ (low_emmission ∨ highlight ∨ smoother ∨ distortion)
If we were to consider the 6 values as logical
propositions we could write:
glass ⇐ similar_color ∧ high_alpha ∧ low_emmission
glass ⇐ similar_color ∧ high_alpha ∧highlight
glass ⇐ similar_color ∧ high_alpha ∧smoother
glass ⇐ similar_color ∧ high_alpha ∧distortion
Multiple Classifiers
We can re-write the previous statement as four
different statements of three propositions:
Multiple Classifiers
Each proposition is a seperatley trained
classifier of lower dimensionCombining the sub-classifiers
Logical OR Weighted Sum Exponential Model
Global Integration
Due to conservativeley built classifiers we will
have few positivesHysteresis: connect positves along a common
edgeSnakes
(Kass, Witkin, Terzopoulos, '87)
Experiments
Training Set: 15 images, 6 with glass objects in
front of various backgrounds, 9 with no glass
objects 333 positive examples 4581 negative examples
Test Set: 50 images, 35 with glass objects, 15
with no glass objects at all
Experiments
Single SVM
Multiple SVM's + OR
Multiple SVM's + Weighted Sum
Multiple SVM's + Exponential Model
Multiple SVM's + Weighted Sum (sampled)
Precision
68.76%
56.04%
58.78%
56.04%
73.70%
Results
Results
Results
Results
Classifying Regions as Glass
We need not restrict ourselves to regions
around edgesGiven two regions we ask the question “is one
region a glass covered version of the other?”
Over Segmentation
We want regions of similar material (Felzenszwalb and Huttenlocher, '04)
Can adjust size of super-pixels (degree of over-
segmentation) with smaller k valuesUse color, texture, and edgels to set weights
Discrepency
We use our previous classifier as a measure of
how much two regions don't belong two the
same material (i.e. glass and not glass) Use distance from seperating hyperplane (Platt, '00)
Large values: far on the postive glass side Small values (negative): far on the not glass side Reasonable if data takes a normal distribution
Drop blur cue since DCT can't be done on non-
rectangular regions.
Ambiguities
Discrepency is high for a material and a glass
covered version of that material, but also for two
completley different materialsAbove example has two possible segmentations
Affinity
Aij = 1 – a
ij /
Affinity
Because of refraction most straight background
edges that pass through the glass will appear
brokenEdges from glass contour ussually the longest
smoothest edges in the area
Affinity
Certainty of Discrepency/Affinity
High discrepency: likely different materialsLow discrepency: cannot ascertain whether
one regions is glass and the other is
backgroundHigh affinity: likely same materialLow affinity: not very informative, edge path
may just have been broken
Objective Function
We wish to maximize our measuresFirst term: maximize discrepency between
glass and other stuffSecond term: maximize affinity in the glassThird term: minimize affinities between glass
and otherCombinatorial problem!
Relaxed Objective Function
Relax region constraints Treat pixels as a sampling of an underlying
continuous function
Geodesic Active Contours
Curve Evolution
Experiments
Single SVM
Multiple SVM's + OR
Multiple SVM's + Weighted Sum
Multiple SVM's + Exponential Model
Multiple SVM's + Weighted Sum (sampled)
Proposed Method
Precision
68.76%
56.04%
58.78%
56.04%
73.70%
77.03%
Results
Results
Results
Results
Results