"How to Create a Great Object Detector," a Presentation from MathWorks

Copyright © 2014 MathWorks, Inc. 1 Your Company Logo

goes here. See Slide

10 for instructions

Avinash Nehemiah

29 May 2014

How to Create a Great Object Detector



10 for instructions

Objectives

• Learn the key steps to create your own object detector

• This does not have to be a 6 month research project

• Tips and Tricks

• How to create a great object detector

• Design considerations for embedded devices



10 for instructions

Why Create Object Detectors ?

Object detectors help us understand real-world scenes.



10 for instructions

What is Object Detection ?

• Detect and locate an object or class of objects in a scene

• Object detection is NOT recognition

• You can’t perform recognition without detection

Specific Object Class of objects

e.g. Faces



10 for instructions

Getting Started with Object Detection

Goal: Locate object and non-object

regions

• Use training data to “learn” what

object looks like

• Learn how to reject non-object

regions

• More compact representation

than image pixels are used

• Known as features

Object Region

Attribution: Examples in this talk created using INRIA Person Dataset

http://pascal.inrialpes.fr/data/human/



10 for instructions

Object Detection is Data Classification

Machine

Learning

Training Data

Features Classifier

Label: ‘Person’/’Non Person’

Response: Likelihood that input

belongs to labelled class

Learned

Classifier



10 for instructions

Creating an Object Detector

Collect Training

Data

Feature Selection

Evaluate Different

Classifiers

Construct Detector

Optimize for Embedded

Implementation

Design Implementation

Attribution: Examples in this talk created using INRIA Person Dataset




10 for instructions


Collect Training

Data

Feature Selection

Evaluate Different

Classifiers

Construct Detector


Implementation




10 for instructions

Training Data

Training Data is Key

• Good sampling of object (positive)

• Different scale, orientation, and

environmental conditions

• Negative data

• Representing environment object

usually appears

• Too few positive samples leads to

insufficient true detection

• Not enough variation in training set

leads to overfitting

• Too few negative samples leads to too

many false positives



10 for instructions

Preparing Data for Training

Label,

Normalize and

Resize

Generate

Negative

Samples

Feature

Extraction

Features #1

Features #2

.

.

.

.

.

.

Features #N

Features #1

Features #2

.

.

.

.

.

.

Features #N

Positive Samples

Negative Samples

Learning Algorithm



10 for instructions


Collect Training

Data

Feature Selection

Evaluate Different

Classifiers

Construct Detector


Implementation




10 for instructions

Feature Extraction

Sparse Dense

SURF HOG Image

Pixels Bag of Words

Feature Extraction

• Representations often invariant to

changes in scale, rotation, illumination

• More compact than storing pixel data

• Feature selection based on nature of

detection problem

• Consider alternatives like convolutional

neural nets

Factors to consider:

• Scale change

• Orientation

differences

• Environmental

conditions



10 for instructions


Collect Training

Data

Feature Selection

Evaluate Different

Classifiers

Construct Detector


Implementation




10 for instructions

Find the Best Classifier-Feature Combination

• Steps

• Use a small subset of training data

• Train several different types of classifiers

• Compare performance

• Cross-validation

• Objective performance metrics



10 for instructions

Important Classification Methods

Supervised Classification

• Support Vector Machines

• E.g. Pedestrian detectors

• Neural Networks

• Great when low variation in orientation and

appearance

Unsupervised Classification

• Gaussian Mixture Model

• Hidden Markov Models

• Works well when large amount of data/classes in

training set

Ensemble Methods

• Adaboost

• E.g. Viola-Jones face detection

• Great when classifier needs to operate in many

different conditions



10 for instructions

Example: Support Vector Machine



10 for instructions

Analyzing Classifier Performance

Option 1: Eye Test

• Look at classification results

Option 2: Receiver Operating Curve

(ROC)

• Means to compare ranges of

performance of classifiers

• Ideal curves are toward the top-

left

• Larger area under the curve is

better

• Do not ignore false alarm rate (

1% of false alarm rate translates

to ~25000/day)



10 for instructions


Collect Training

Data

Feature Selection

Evaluate Different

Classifiers

Construct Detector


Implementation




10 for instructions

Sliding

Window

Object Classifier Label/

Response

Constructing the Detector

Label: ‘Person’/’Non Person’

Response: Distance from boundary



10 for instructions

Constructing the Detector

• Classify patches at each

location and scale

• Resolve detection scores

• Pick strongest bounding box



10 for instructions


Collect Training

Data

Feature Selection

Evaluate Different

Classifiers

Construct Detector


Implementation




10 for instructions

Considerations for Embedded Vision

Memory

• Memory footprint of classifier model

Processing Time

• Classification time for each window-step

• Granularity of sliding window step size

• Cascading-classifiers for faster rejection of negatives

Potential Porting Issues

• Fixed point accuracy

Tip: Consider dimensionality reduction

E.g. For pedestrian detector 99.3% speedup using PCA dimensionality reduction



10 for instructions

Classifier Training Workflow

Collect Training Data

Feature Selection

Evaluate Different

Classifiers

Construct Detector


Implementation

Few iterations might be necessary to create a GREAT detector

Computer Vision System Toolbox

Parallel Computing Toolbox

Statistics Toolbox

MATLAB Coder ( C/C++)

HDL Coder

Fixed Point Designer



10 for instructions

Key Takeaways

• It is not difficult to create your own object detector

• Smart feature and classifier selection helps create great object

detectors

• Never ever rely on solely the “eye test” for evaluation

• Insufficient training data will hold you back



10 for instructions

Resources

Computer Vision and Machine Learning with MATLAB

• http://www.mathworks.com/products/computer-vision/

• http://www.mathworks.com/machine-learning/index.html

INRIA Person Dataset

• http://pascal.inrialpes.fr/data/human/

Presentations

• http://courses.engr.illinois.edu/cs543/sp2011/lectures/Lecture%20

19%20-%20Sliding%20Window%20Detection%20-

%20Vision_Spring2011.pdf

Contact [email protected] for more information or source code

http://www.mathworks.com/products/computer-vision/



http://www.mathworks.com/machine-learning/index.html




http://courses.engr.illinois.edu/cs543/sp2011/lectures/Lecture 19 - Sliding Window Detection - Vision_Spring2011.pdf






mailto:[email protected]

"How to Create a Great Object Detector," a Presentation from MathWorks

Technology

instructions object

object detectors

company logo

environment object

object detection goal

scene object detection

data training data

nonobject regions