Professional Documents
Culture Documents
-Microsoft Research
,
, 2010
?
?
/
, ,
Enzweiler & Gavrila. Monocular Pedestrian Detection: Survey and Experiments. TPAMI09
Enzweiler & Gavrila. Monocular Pedestrian Detection: Survey and Experiments. TPAMI09
1.
2.
3. :
/range maps
4.
5.
( , ...)
Gould, Fulton, Koller, ICCV09
:
-
-
-
:
-
Disclaimers
( !)
:
Pictorial structures
Generative models
( online):
Li Fei-Fei, R. Fergus, A.Torralba,
Recognizing and Learning
Object Categories,
ICCV05, CVPR07, ICCV09
B. Leibe, K. Grauman,
Visual Object Recognition, AAAI08
Dataset
(Weizmann Horses)
:
1 =
0 =
Daimler pedestrians data
:
1.
2.
1.
2.
3. Branch-and-Bound
1:
f()
(HOG)
Dalal & Triggs, CVPR 2005
1.
2.
3.
4.
0 45 90135
0 45 90135
0 45 90135
0 45 90135
: SIFT, SURF, DAISY, etc.
: +1 -1
SVM :
:
[Cortes&Vapnik, ML93]
loss
1
0
-
Dalal&Triggs, CVPR 2005: HOG+ SVM
W+
W-
-
( ):
+
-
>
1 ?
+ - + >
+ >
- +
2 ?
3 ?
loss
1
0
:
:
loss
1
0
:
1.
()
2.
3.
4.
-
[Viola & Jones, CVPR01]
1- 2- ,
?
A = (A+B+C+D) (B+C) (B+D) + B
- vs. -
-
- ( )
[Dollar et al. BMVC09]: -
-
[Sivic & Zisserman, ICCV03]
[Csurka et al. ECCVw 04]
[Sivic & Zisserman, ICCV03]
[Csurka et al. ECCVw 04]
: , , ,
Sivic & Zisserman, ICCV 2003
Sivic & Zisserman, Video Google, ICCV 2003
:
http://www.robots.ox.ac.uk/~vgg/research/
vgoogle/
:
SVM
(w)
SVM
SVM:
, SVM Branch-and-Bound
[Lampert , Blaschko, Hofmann CVPR08]
SVM (w)
(w h(
))
(w+ h(
)) + (w- h(
))
[Lazebnik, Schmid & Ponce, CVPR 2006]
Branch-and-Bound
[Hough1959], [Duda & Hart, Comm. ACM 1972]
(-)
[Lowe ICCV99]
[Ballard PR81]
[Lowe ICCV99]
[Leibe & Schiele DAGM04]
[Leibe, Leonardis & Schiele IJCV08]
+
y
s
y
[Leibe, Leonardis & Schiele IJCV08]
s
x
-
Juergen Gall & Victor Lempitsky,
Class-Specific Hough Forests for Object Detection, CVPR 2009
Jrgen Gall
[Breiman et al.84]
3
f1[x y 1] T 0
4
f2[x y 1]T 0
1
2
6
1
1. Yali Amit, Donald Geman: Shape quantization and recognition with randomized trees.
Neural Computation, 1997.
2. Leo Breiman: Random forests. Machine Learning, 2001.
-
[Breiman JMLR01]
function Node = _( {(x,y)}, Level)
{
if {y} Level == maxLevel
return _( y);
{tests} = _N__({(x,y)},N);
test = ___({tests});
{(x0,y0)} = {(x,y) | test(x) = 0};
{(x1,y1)} = {(x,y) | test(x) = 1};
LeftChild = _( {(x0,y0)}, Level+1);
RightChild = _( {(x1,y1)}, Level+1);
return _(test, LeftChild, RightChild);
}
//
function main() {
{X,Y} = __();
for i = 1 to N
{Xi,Yi} = _({X,Y}));
TreeRoot_i = _({Xi,Yi});
end
}
E(S) = 0
:
(information gain):
E(S) = 1
-:
I(p) I(q) +
p
q
30 (, , HOG-like)
[Moosmann et al. NIPS 06], [Winn & Shotton CVPR 06]
-:
:
+
-:
:
+
:
-:
-:
55
-:
-:
-:
-:
, 6
-:
:
60
-:
()
:
(100.000+ )
[Desai, Ramanan & Fowlkes. Discriminative models for multi-class object layout,
ICCV09 best paper]