Graz University of Technology, AUSTRIA Institute for Computer Graphics and Vision Fast Visual Object Identification and Categorization Michael Grabner, Helmut Grabner, Horst Bischof
Dec 19, 2015
Graz University of Technology, AUSTRIA
Institute for Computer Graphics and Vision
Fast Visual Object Identification and Categorization
Michael Grabner, Helmut Grabner, Horst Bischof
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 2 (of 19)
Agenda
Motivation
Approach
Experimental Illustration
Results
Outlook
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 3 (of 19)
Problem
Database: Ferencz, Yale, Buffalo
How large scale object recognition can be handled in an adequate time?
How knowledge can be used for incremental learning from few examples?
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 4 (of 19)
Identification vs. Categorization
Faces
Writings
Cars
Horst boring Joe wondering
Bill‘s carZip Code 77840
Horst laughing
Identification Categorization
. . .
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 5 (of 19)
Identification and Categorization
Faces
Horst
Helmut
Joe
Cars
Car 1
Car 2
Car 3
Car 4
Writings
ZIP Codes
Places
wondering
Identification depends on the granularity of categorization
tired
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 6 (of 19)
Our approach
„Object Memory“- Hierarchical meaning objects are stored in a hierarchical way
- Incremental meaning objects can be added incrementally to the structure
- Fast meaning identification of objects is done efficiently
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 7 (of 19)
Features
Two types of features- Haar-Like (Viola and Jones 2001)
- Orientation Histograms
Advantages- Coding of gradient information (Lowe 2004, Edelman 1997)
- Fast computation allows to extract a large number of features leading to robustness (Porikli 2005, Grabner 2005)
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 8 (of 19)
Integral Orientation Histogram
F. Porikli: „Integral histograms: A fast way to extract histograms in Cartesian spaces“, in Proc. CVPR 2005
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 9 (of 19)
Feature Selection
Goal is to distinguish between objects by selecting discriminative features
Feature Pool Learn distance function (Ferencz 2005)
- „same“ vs. „same“ and „same“ vs. „different“
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 10 (of 19)
1.) A weak classifier corresponds to a single feature
2.) Perform boosting to select N features
3.) Final strong classifier is a linear combination of features
Boosting for Feature Selection (Viola and Jones 2001)
selected FeaturesObject model
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 11 (of 19)
Building the „Object Memory“
Initialization: 2 objects form a single layer
Adding a novel object:
- Evaluating the sample starting at the highest layer• If sample can not be modeled by one of the classifiers: ADD TO
CURRENT LAYER
• If sample can be modeled by one of the classifiers: GO DEEPER– If classifier has no child: INITIALIZE A NEW LAYER
Retrain- current layer to distinguish between these models- parents for getting generic object models in higher layers
Generating layers of similar objects and learn to differentiate between these similar objects
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 12 (of 19)
Building the „Object Memory“
Training the Object Memory
On-line Illustration MATLAB
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 13 (of 19)
Identification Process
Evaluating the sample starting at the highest level
Multi-path evaluation based on model confidences
Post Processing (i.e. take reference model with highest confidence)
Note: evaluation is fast using integral data structures
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 14 (of 19)
Identification Process
Evaluation the Object Memory
On-line Illustration MATLAB
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 15 (of 19)
Experiments - Overview
Experiment 1- Illustration of the approach
- 3 categories (Cars, Faces, Writings)
- Training using 6 images per object
- Model complexity: 30 features
Experiment 2- Performance evaluation on category Cars
- Varying number of objects and model complexity
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 16 (of 19)
Experiment 1 – Trained Object Memory
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 17 (of 19)
Experiment 2
Experiment on database Car (Ferencz)
- 6 samples for training (const)
- RPC obtained by varying confidence threshold
Variation of model complexity (30 Objects) Variation of objects (15 Features)
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 18 (of 19)
Conclusion and Outlook
Conclusion- Hierarchical structuring of objects by a simple heuristic
- Incremental adding of novel objects from few examples
- Fast Identification
Outlook- More objects
- Fast and efficient retraining• On-line boosting for model update
- Detection, Tracking and Recognition within one framework• all tasks are performed with same types of features
NIPS 2005 Workshop: Interclass Transfer„why learning to recognize many objects is easier than learning to recognize just one“
Slide 19 (of 19)
Thank you for your attention!