Slides - ETHZ - Computer Vision Lab
Slides - ETHZ - Computer Vision Lab
Slides - ETHZ - Computer Vision Lab
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Learning for Tracking andLessons Learned from itHelmut Grabner
Tracking CuesObject AppearanceBackgroundObject/BackgrounddiscriminationObject BoundaryMotion2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>4
[Grabner et al. VideoProc.CVPR 2006]2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>5
Tracking Requirements (model free tracking) Adaptive RobustnessAnd of course,REAL TIME! Generality2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>6
PART IOn-line Boosting basedTrackingCVPR’06, BMVC’06
Boosting and <strong>Vision</strong> Boosting[Freund and Schapire, JCSC, 1997] On-line boosting[Oza and Russel, AIS, 2001] Boosting for FeatureSelection[Tieu and Viola. CVPR 2000], [Violaand Jones, CVPR 2001] On-line Boosting forFeature Selection[Grabner and Bischof, CVPR 2006]2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>8
Off-line learning2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>9
On-line learning<strong>Lab</strong>eled InformationLearning AlgorithmTeacher2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>10
Sir, could you please tellme what boosting is?Boosting needs just someguy who is a little bit betterthan guessing.2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>11
Off-line boostingStrong classifierWeak classifierReweighting thetraining examples2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>12
Boosting for Feature Selection Combination of SimpleImage Features fordistinguishing two classes Features = weak classifier Boosting to select a subset(strong classifier)2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>13
On-line Boosting2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>14
On-line BoostingConverges to theoff-line result!2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>15
On-Boosting for Feature Selection Combination of SimpleImage Features fordistinguishing two classes Features = weak classifier Boosting to select a subset(strong classifier)2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>16
On-Boosting for Feature Selection Combination of SimpleImage Features fordistinguishing two classesGeneral approachfor on-line featureselection. Features = weak classifier Boosting to select a subset(strong classifier)2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>17
Tracking as Classificationobjectvs.background2009/08/18, 2008/11/06 Scovis Southampton Meeting H. Department Grabner, Tracking of Information Learning Technology and Lessons and Electrical Learned Engineering from it - ETH-Zurich, – <strong>Computer</strong> <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>oratory <strong>Vision</strong> 18
Tracking as Classificationobjectvs.background2009/08/18, 2008/11/06 Scovis Southampton Meeting H. Department Grabner, Tracking of Information Learning Technology and Lessons and Electrical Learned Engineering from it - ETH-Zurich, – <strong>Computer</strong> <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>oratory <strong>Vision</strong> 19
from time t to t+1evaluate classifier on sub-patchesactual object positionsearch Regionanalyze map and set newobject positionupdate classifier (tracker)create confidence map--+- -2009/08/18, 2008/11/06 Scovis Southampton Meeting H. Department Grabner, Tracking of Information Learning Technology and Lessons and Electrical Learned Engineering from it - ETH-Zurich, – <strong>Computer</strong> <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>oratory <strong>Vision</strong> 20
ObjectDetectorFixed Training setGeneral object detectorOff-line Boosting forFeature SelectionObject TrackerOn-line updateObject vs. BackgroundOn-Line Boosting forFeature Selection2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>21
LESSON LEARNED 1Tracking is a simple task!(When formulating it properly)
2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>23
“Tracking the Invisible”2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>25
Tracking Solved ☺Does it fail?If yes, when?2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>26
When does it fail…2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>27
When does it fail…Often, all toooften!2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>28
When does it fail…WHY?2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>29
from time t to t+1evaluate classifier on sub-patchesactual object positionsearch Regionupdate classifier (tracker)-+-analyze map andset new objectpositionSelf-learningcreate confidence map- -2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>30
Drifting due to self-learning policyTracked PatchesConfidence2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>31
LESSON LEARNED 2Self-learning → drifting!
PART IISemi-Supervised On-lineBoosting for TrackingECCV’08
Review: Supervised Tracking2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>34
Problems of…-<strong>Lab</strong>el Jitter+++<strong>Lab</strong>el Noise2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>35
Semi-Supervised TrackingUn-labeleddata<strong>Lab</strong>eled data2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>36
Supervised learning+ -+ -Maximum margin2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>37
Can Unlabeled Data Help??? ?? ? ???? ? ?? ?? +? ?-??? ? ???? ?? ? +? ? - ????? ? ? ? ?decision? ?? ?boundary in lowdensity region2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>38
Semi-Supervised On-line BoostingPrior+ + +???-????2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>39
Semi-Supervised On-line BoostingPrior“stable”fixdynamicNobody is perfect!But, be a honest Teacher!H off (Prior) can be wrong with lowconfidence.2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>40
Tracking LoopPrioractual object positionevaluate classifier onsub-patchessearch Regionupdate classifier (tracker)create confidence map????2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>41
Prior<strong>Lab</strong>eled dataUn-labeleddata2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>42
LESSON LEARNED 3On-line Semi-supervisedlearninig → limiteddrifting.
Occlusions2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>44
Object disapearance2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>45
Long term tracking (1h)Click here to start2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>46
Tracking Solved ☺Does it fail?If yes, when?2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>47
2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>48
Prior is toogeneric (e.g, dirftto similar objects)2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>49
2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>50
Prior restricts toomuch (e.g., partialocclusions)2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>51
LESSON LEARNED 4Prior is essential in semi-supervised learning.(c.f., Stability PlasticyDilemma)
PART IIIBeyond Semi-SupervisedTrackingICCV’09 WS on On-line Learning for<strong>Computer</strong> <strong>Vision</strong>
Review: DetectionNon adaptive at all!2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>54
Review: Supervised TrackingToo adaptive2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>55
Review: Semi-Supervised TrackingLimit drifting but toorestrictive / general2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>56
NEW PRINCIPLE: Active Sampling via Tracking2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>57
NEW PRINCIPLE: Active Sampling via Tracking2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>58
NEW APPROACH: Adaptive PriorAdditionalinformationfor prior updateDetectorDetector2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>59
Specializing (Simplifying)Updates No Few trustfulLearning taskApplicableObject class vs.everything elseAny time,everywhereParticular objectsvs. other objectsand backgroundCurrent sceneMany semisupervisedCurrent Objectappearance vs.local surroundingLocalneighborhood2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>60
Multiple classifier system#include “vision.h” use addition information, e.g., multipleobjects, background imageInformation aggregation2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>61
2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>62
2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>63
Performance evaluationpositionTruepositivesTooAdaptiveNotvisibleFalsepositivesFalsenegativesLessAdaptiveGround Truthtime2009/03/23 2009/08/18, Oxford Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>64
Implicit Occlusion handling2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>65
Really Long Term Tracking (24h)Click here to start2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>66
……2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>67
LESSON LEARNED 5<strong>Vision</strong> is more then pureMachine Learning!(keep problems simple)
Extension 1: re-IdentificationIdentifier 1Identifier 22009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>69
2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>70
Extension 2: Information AggregationWhole ImageScene / Current object /local surroundingSpecializationObject instance /current sceneSpecific imagelocation2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>71
Pedestrian Detection (PETS)Generic Detector & ContextProposed Approach2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>72
LESSON LEARNED 6<strong>Vision</strong> ≠ Detection +Tracking + Recognition(benefitting from a lot of– unlabeled – data)2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>73
ConclusionTracking issimpleSelf-learning →driftingSemi-SupervisedlearningPrior <strong>Vision</strong> > MachineLearning<strong>Vision</strong> ≠ Det. +Trac. + Rec.2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>74
AcknowledgmentsLuc van GoolSeverin StalderEU-project SCOVIS under grantagreement no 216465.Institute for <strong>Computer</strong> Graphics and <strong>Vision</strong>Graz, University of Technology, AustriaCentere of Machine PreceptionCzech Technical UniversityHorst BischofMichael GrabnerChristian LeistnerJiri MatasJan Sochamn2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>75
Code/Demos & Tracker Evaluation2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>76
Additional <strong>Slides</strong>…2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>77
DetectorNo UpdatesValid samplesConfidence map2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>78
Detectorpos.updatesValid samplesConfidence mapNeg.updates2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>79
DetectorValid samplesConfidence mapUnlabeled updates(foreground & local background)2009/08/18, Southampton H. Grabner, Tracking for Learning and Lessons Learned from it - ETH-Zurich, <strong>Computer</strong> <strong>Vision</strong> <strong>Lab</strong>80