150 likes | 261 Vues
This paper presents an evaluation of manually created ground truth data for multi-view people localization. We define the problem of accurate localization using multiple cameras in overlapping fields of view and the necessity for reliable ground truth. Our framework incorporates a GUI for real-time triangulation and localization of individuals using feedback mechanisms. We analyze the performance of human observers in generating ground truth, addressing issues like errors and recall rates. We emphasize the importance of precise ground truth for valid algorithm evaluation and suggest automation to enhance reliability.
E N D
Evaluation of ManuallyCreatedGroundTruthforMulti-viewPeopleLocalization Ákos Kiss, Tamás Szirányi DistributedEventsAnalysis Research Laboratory kiss.akos@sztaki.mta.hu Sziranyi.tamas@sztaki.mta.hu
Multi-ViewPeopleLocalization Overview • Problemdefinition • Motivation • Ourframework • Contribution • Referencegroundtruth • Evaluating human observers • Summary
Multi-ViewPeopleLocalization Problemdefinition • Multiplecameras monitoring an area • Overlappingfield of view • Additionalspatialinformation • Pixelscorrespondtolinesinspace • Linesintersectwhereobjectlies • Triangulation • Groundtruthrequired
Multi-ViewPeopleLocalization Problemdefinition • Creatinggroundtruth • Manually • Positioninimages • Boundingboxes • Positioninreferencespace • Parametricsurface • True 3D positions • Automatically • ToFsensors • Kinect • Lidar - Expensivesolutions - Occlusion is still a problem
Multi-ViewPeopleLocalization Motivation • Groundtruthusuallytakenforgranted • Human makemistakes • Poorgroundtruthleadstoinvalidalgorithmevaluation • Generatinggroundtruth is timeconsuming • Experts’ time is expensive • Evaluating human observers • 9 subjects (6 laymen, 3 withdomainknowledge)
Multi-ViewPeopleLocalization Ourframework • GUI – allviewsvisible • Locationbytriangulation • Mark inany (atleast 2) views • Triangulation • Feedback (savelocationonlyifcorrect) • Localizingpersonbyperson • Head • Feet • Skipfootifnotvisiblein 2 views
Multi-ViewPeopleLocalization Ourframework • Localizingbylinearoptimization • p is onbothlines ( axis): • Linespracticallyneverintersect • Pseudoinverse (minimalizes SSE) • Validatingwithknownsurface • Planarground is reconstructedprecisely
Multi-ViewPeopleLocalization Referencegroundtruth • Several „groundtruth” createdbysubjects • Combininginformationfrommultiplegroundtruth • Peoplearelocalizedbyonly a subset of subjects • Positionsarenoisy • Automatingprocess • Match peoplebyhead (body) location • Matchingfeet (ordermightdiffer) • Filteringnoise (weighted center of locations) • More reliableresult: referencegroundtruth
Multi-ViewPeopleLocalization Evaluating human observers • Errors (precision) • Accuracy (locationerror) • Recall (missedpeople) • Temporalanalysis
Multi-ViewPeopleLocalization Evaluating human observers • Errors (precision) • Lowerrorrate • Typicalerrors • Parallel lines (mark near camera) • Mix uppeople • Accuracy (locationerror) • Recall (missedpeople) • Temporalanalysis
Multi-ViewPeopleLocalization Evaluating human observers • Errors (precision) • Accuracy (locationerror) • Synchronizationerror • Lowdeviation • Outliersuppression • Iterative • Changeweight of points • Recall (missedpeople) • Temporalanalysis
Multi-ViewPeopleLocalization Evaluating human observers • Errors (precision) • Accuracy (locationerror) • Recall (missedpeople) • Verylowrecall • Expertsarenotbetter • Temporalanalysis feet = headrecall line recallvalues of laymen (blue) and experts (green)
Multi-ViewPeopleLocalization Evaluating human observers • Errors (precision) • Accuracy (locationerror) • Recall (missedpeople) • Temporalanalysis • Shortexperiment • Fewsubjects • Less outliersforexperts
Multi-ViewPeopleLocalization Summary • Typical human observer • Highprecision • Highaccuracy • Lowrecall • Generatingthistype of groundtruthrequiresmuchattention • Generating more reliablereferencegroundtruthcan be automated