Xuehan Xiong Daniel Munoz Drew Bagnell Martial Hebert 1 2 Problem 3D Scene Understanding C ar P ole G round T runk W ire B uilding V eg 3 Solution Contextual C lassification ID: 399118
Download Presentation The PPT/PDF document "3-D Scene Analysis via Sequenced Predict..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
3-D Scene Analysis via Sequenced Predictions over Points and Regions
Xuehan Xiong
Daniel Munoz
Drew Bagnell
Martial Hebert
1Slide2
2
Problem: 3D Scene Understanding Slide3
C
ar
Pole
Ground
T
runk
W
ire
B
uilding
V
eg
3
Solution: Contextual
C
lassificationSlide4
Intractable inference
Difficult to train
Limited success
4
Graphical models
Fig. from Anguelov
, et al. CVPR 2005
Classical Approach: Graphical
M
odels
Anguelov
, et al. CVPR
2005
Triebel
, et. al. IJCAI 2007
Munoz, et al. CVPR
2009
Kulesza
NIPS 2007Wainwright JMLR 2006
Finley &
Joachims
ICML 2008
Belief propagation
Mean field
MCMCSlide5
Intractable inference
Difficult to train
Limited success
5
Graphical models
Fig. from
Anguelov
, et al. CVPR 2005
Classical Approach: Graphical
M
odels
Anguelov
, et al. CVPR
2005
Triebel
, et. al. IJCAI 2007
Munoz, et al. CVPR
2009
Kulesza NIPS 2007
Wainwright JMLR 2006
Finley &
Joachims
ICML 2008
Belief propagation
Mean field
MCMCSlide6
Intractable inference
Difficult to train
Limited success
6
Graphical models
Fig. from Anguelov, et al. CVPR 2005
Classical Approach: Graphical
M
odels
Anguelov
, et al. CVPR
2005
Triebel
, et. al. IJCAI 2007
Munoz, et al. CVPR
2009
Kulesza
WainwrightFinley &
Joachims ICML 2008
Belief propagation
Mean field
MCMCSlide7
7Slide8
8Slide9
9Slide10
10Slide11
Our Approach: Inference Machines
11
Train
an
inference
procedure
, not a model.
To encode spatial layout and long range relations
Daume
III
2006,
Tu
2008,
Bagnell
2010, Munoz 2010Slide12
Train an inference
procedure
, not a model.
To encode spatial layout and long range relations
Daume
III 2006,
Tu
2008,
Bagnell
2010, Munoz
2010
Inference via sequential prediction
12
C0
C1
C2
Reject
T
T
F
F
F
E.g. Viola-Jones 2001
Our Approach: Inference MachinesSlide13
13
C0
C1
C2
context
context
O
urs
Train an inference
procedure
, not a model.
To encode spatial layout and long range relations
Daume
III 2006,
Tu
2008,
Bagnell
2010, Munoz
2010
Inference via sequential prediction
Our Approach: Inference MachinesSlide14
p
oint features
14
Example featuresSlide15
point features
15Slide16
point features
16
t
op
mid
bottom
Contextual featuresSlide17
point features
17
t
op
mid
bottomSlide18
point features
18
t
op
mid
bottomSlide19
point features
19
t
op
mid
bottomSlide20
point features
20
t
op
mid
bottomSlide21
Local features only
21
C
ar
Pole
B
uilding
V
eg
G
round
W
ireSlide22
Round 1
22Slide23
Round 2
23Slide24
Round 3
24
C
ar
V
egSlide25
Create regions
Level 2
Level 1
25Slide26
26Slide27
region features
27
Region level
Pt
levelSlide28
Level 2
Level 1
28Slide29
point features
29
P
oint level
Region levelSlide30
With Regions
30Slide31
Learned Relationships
31
Neighbor contextual feature
Learned weights
point features
t
op
mid
bottomSlide32
Learned Relationships
32
Neighbor contextual feature
Learned weights
point features
t
op
mid
bottomSlide33
Experiments
3 large-scale datasetsCMU (26M), Moscow State (10M), Univ. Wash (10M)Multiple classes (4 to 8)
car, building, veg, wire, fence, people, trunk, pole, ground, street signDifferent sensors
SICK (ground), ALTM 2050 (aerial), Velodyne (ground)Comparisons
Graphical models, exemplar based33Slide34
Quantitative Results
34
[1] Munoz CVPR 2009
[2]
Shapovalov
PCV 2010
[3] Lai
RSS 2010 *
* Use additional semi-supervised data
not leveraged
by other methods.Slide35
CMU Dataset
Ours
Max Margin CRF [1]35
[1] Munoz, et. al.
CVPR 2009Slide36
Ours
Max Margin CRF [1]
36
CMU Dataset[1] Munoz, et. a
l. CVPR 2009Slide37
Ours
Max Margin CRF [1]
37
CMU Dataset[1] Munoz, et. a
l. CVPR 2009Slide38
Moscow State Dataset
Ours
Logistic regression
38Slide39
Conclusion
Simple and fast approach for scene labelingNo graphical modelLabeling via 5x logistic regression predictions
Support flexible contextual featuresLearning rich relationships
39
C0
C1
C2
context
contextSlide40
Thank you! And Questions?
AcknowledgementsUS Army Research Laboratory, Collaborative Technology AllianceQinetiQ North America Robotics Fellowship
40