SLIDE 1
Marius Leordeanu, Martial Hebert, Raul Sukthankar CVPR07 CVPR 07 - - PowerPoint PPT Presentation
Marius Leordeanu, Martial Hebert, Raul Sukthankar CVPR07 CVPR 07 - - PowerPoint PPT Presentation
Marius Leordeanu, Martial Hebert, Raul Sukthankar CVPR07 CVPR 07 Presented by Weina Ge Problem Problem Image categorization and localization , given negative images and weakly labeled positive images Contributions Learning
SLIDE 2
SLIDE 3
SLIDE 4
Features Localization as feature matching
g
Matching score: Quadratic assignment problem (QAP): Quadratic assignment problem (QAP):
▪ Solved by spectral matching algorithm with one‐to‐one mapping constraints
SLIDE 5
h ll b h
There will be an assignment in matching no matter
what, so we need to do recognition
Model the posterior
Model the posterior:
▪ C = 1 if the object is present at location x*; otherwise C=0 ▪ D – data
Posterior is a function of
▪ quality of matching ▪ quality of matching ▪ quality of model parts: relevant to the particular category and
discriminative against the negative class
SLIDE 6
Approximate the posterior by a logistic classifier
pp p y g
1: detected 0: absent
- matching quality (pairwise potentials)
0 abse
- model quality (relevance parameter )
- sigmoid function
h h l h
- squashes the relevance parameter to either 1 or 0
- reduces the overfitting without an explicit regularization term
SLIDE 7
d l
Model parameters
Model parts and their geometric relationships
=
Sensitivity to geometric deformations
= Learned independently of object parts and object class
Relevance parameter
Sequential gradient descent
Objective function
SLIDE 8
Each model usually contains 40~100 parts
y p
SLIDE 9
M i f t f diff t i i t
Merging features from different viewpoints
M d l t t tl b i dd d d d
Model parts are constantly being added and removed
SLIDE 10
L i i l ifi d l h i h i i f
Logistic classifiers model the posterior that a given pair of
assignments is correct given the geometric deformation
Manually‐selected correspondences
8000 pairs of correct ones and 16000 pairs of wrong ones per database 3 databases: CALTECH 5 INRIA horses GRAZ 02 3 databases: CALTECH‐5, INRIA‐horses, GRAZ‐02
Experiments imply the space of geometric second‐order
d f i i l h dl f bj deformations is more or less the same regardless of object classes vs
SLIDE 11
SLIDE 12
PASCAL dataset (587 images)
PASCAL dataset (587 images)
“Ours” VS Winn et al. (different features)
geometric constraints VS appearance
geometric constraints VS appearance
SLIDE 13
d
GRAZ datasets
“Ours” VS Opelt et al. (different feature selection criteria)
▪ performance of a group of features VS individual features ▪ performance of a group of features VS individual features
SLIDE 14