Visual tracking with online multiple instance learning
This presentation is the property of its rightful owner.
Sponsored Links
1 / 23

Visual Tracking with Online Multiple Instance Learning PowerPoint PPT Presentation


  • 402 Views
  • Uploaded on
  • Presentation posted in: General

Visual Tracking with Online Multiple Instance Learning. 2014-02-04 Ko Daewon. Boris Babenko , Ming- Hsuan Yang, Serge Belongie CVPR 2009. Visual Tracking with Online Multiple Instance Learning. C ontent. Introduction Tracking by Detection Multiple Instance Learning (MIL)

Download Presentation

Visual Tracking with Online Multiple Instance Learning

An Image/Link below is provided (as is) to download presentation

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.


- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -

Presentation Transcript


Visual tracking with online multiple instance learning

Visual Tracking with Online Multiple Instance Learning

2014-02-04

KoDaewon

  • Boris Babenko, Ming-Hsuan Yang, Serge Belongie

  • CVPR 2009


C ontent

Visual Tracking with Online Multiple Instance Learning

Content

  • Introduction

  • Tracking by Detection

  • Multiple Instance Learning (MIL)

  • Online MILboost

  • Experiments

  • Conclusion


Introduction

Visual Tracking with Online Multiple Instance Learning

Introduction

  • Problem: Track a object in video in given location in first frame

  • Typical Tracking System:

    • Appearance Model

      • Color , subspaces, feature, etc

    • Optimization/Search

      • Greedy local search, etc


System overview miltrack

Visual Tracking with Online Multiple Instance Learning

System Overview : MILTrack


Tracking by detection

Visual Tracking with Online Multiple Instance Learning

Tracking by Detection

  • Recent tracking work

    • Focus on appearancemodel

    • Borrow techniques from object detection

      • Slide a discriminative classifier around image


Tracking by detection1

Visual Tracking with Online Multiple Instance Learning

Tracking by Detection

  • Online AdaBoost :

  • Grab one positive patch, and some negative patch, and train/update the model.

negative

positive

Classifier

Online Classifier(Online AdaBoost )


Tracking by detection2

Visual Tracking with Online Multiple Instance Learning

Tracking by Detection

  • An illustration of how most tracking by detection systems work

Probablity map

Frame t+1

Frame t

Frame t+1

X

X

old location

new location

negative

positive

Model

Model

Step 1: Update Appearance Model

Step 2: Update Appearance Model inside of window around old location

Step 3: Update

Tracker state


Tracking by detection3

Visual Tracking with Online Multiple Instance Learning

Tracking by Detection

  • Repeat:

negative

positive

negative

positive

Classifier

Classifier


Tracking by detection4

Visual Tracking with Online Multiple Instance Learning

Tracking by Detection

  • Problems:

  • What if classifier is a bit off?

    • Tracker starts to drift

  • How to choose training examples?


Multiple instance learning mil

Visual Tracking with Online Multiple Instance Learning

Multiple Instance Learning(MIL)

Instead of instance, get bagof instances

Bag is positive if one or more of it’s members is positive

A set of image patches

:Positive

:Negative


Multiple instance learning mil1

Visual Tracking with Online Multiple Instance Learning

Multiple Instance Learning(MIL)

Updating a discriminative appearance model:

(A)

(B)

(C)

MIL Classifier

Classifier

Classifier


Multiple instance learning mil2

Visual Tracking with Online Multiple Instance Learning

Multiple Instance Learning(MIL)

  • MIL Training Input

  • The bag labels are defined as:


Online milboost

Visual Tracking with Online Multiple Instance Learning

Online MILBoost

Framet

Framet+1

Get data (bags)

Update all M classifiers

in pool

Greedily add best K classifiersto

strong classifier


Boosting

Visual Tracking with Online Multiple Instance Learning

Boosting

  • Train classifier of the form:

  • where is a weak classifier

  • Can make binary predictions using


Online milboost1

Visual Tracking with Online Multiple Instance Learning

Online MILBoost

At tframe, Update all Mcandidate weak classifiers

Pick best Kin a greedy fashion (M>K)


Online milboost2

Visual Tracking with Online Multiple Instance Learning

Online MILBoost

  • When the weak classifier receives new data

  • Use update rules:

  • The update rules for and are

  • similarly defined


Online milboost3

Visual Tracking with Online Multiple Instance Learning

Online MILBoost

  • Objective to maximize: Log likelihood of bags:

    where:

Noisy-OR Model, The bag probability

The instance probability


Online milboost4

Visual Tracking with Online Multiple Instance Learning

Online MILBoost

M>K,

M : total weak

classifier candidates

K : choosing the

best K classifiers


Online milboost vs online adaboost

Visual Tracking with Online Multiple Instance Learning

Online MILBoostvs Online AdaBoost


Experiments

Visual Tracking with Online Multiple Instance Learning

Experiments

  • OAB:OnlineAdaBoost

  • SemiBoost:Online Semi-supervised Boosting

  • FragTrack= Stactic appearance model


Experiments1

Visual Tracking with Online Multiple Instance Learning

Experiments


Experiments2

Visual Tracking with Online Multiple Instance Learning

Experiments


Conclusion

Visual Tracking with Online Multiple Instance Learning

Conclusion

  • Present MILTrack that uses a novel Online Multiple Instance Learning algorithm

  • Using MIL to train an appearance model results in more robust tracking


  • Login