Marius Leordeanu, Martial Hebert, Raul Sukthankar CVPR07 CVPR 07 - PowerPoint PPT Presentation

Marius Leordeanu, Martial Hebert, Raul Sukthankar CVPR’07 CVPR 07 Presented by Weina Ge

� Problem � Problem � Image categorization and localization , given negative images and weakly labeled positive images � Contributions � Learning object/category models in a weakly ‐ supervised L i bj t/ t d l i kl i d shape fashion ▪ Object models: cliques of fully ‐ interconnected parts ▪ Pairwise geometric relationship � Simple features ▪ Sparse points and their normals ▪ Sparse points and their normals

� 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

� There will be an assignment in matching no matter h ll b h 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

� Approximate the posterior by a logistic classifier pp p y g 1: detected 0: absent 0 abse matching quality (pairwise potentials) � model quality (relevance parameter ) � sigmoid function � squashes the relevance parameter to either 1 or 0 h h l h � reduces the overfitting without an explicit regularization term �

� Model parameters d l � 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

� Each model usually contains 40~100 parts y p

� Merging features from different viewpoints M i f t f diff t i i t � Model parts are constantly being added and removed M d l t t tl b i dd d d d

� Logistic classifiers model the posterior that a given pair of L i i l ifi d l h i h i i f 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 deformations is more or less the same regardless of object i i l h dl f bj classes vs

� PASCAL dataset (587 images) PASCAL dataset (587 images) � “Ours” VS Winn et al. (different features) � geometric constraints VS appearance geometric constraints VS appearance

� GRAZ datasets d � “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

� Clear theme l h � Proof shape is an important cue for recognition � Careful design of experiments � Every set of experiments serves a purpose Every set of experiments serves a purpose