Transcription of Histograms of Oriented Gradients for Human Detection - Inria
{{id}} {{{paragraph}}}
HistogramsofOrientedGradientsforHumanDet ectionNavneetDalalandBillTriggsINRIARh studythequestionoffeature setsforrobustvisualob-jectrecognition,ad optinglinearSVMbasedhumandetec-tionasa testcase. Afterreviewingexistingedge andgra-dientbaseddescriptors, weshowexperimentallythatgridsofHistogram sofOrientedGradient(HOG)descriptors sig-ni cantlyoutperformexistingfeature studythein uenceofeach stage ofthecomputationonperformance, concludingthat ne-scalegradients, neorientationbinning, relativelycoarsespatialbinning, andhigh-qualitylocalcontrastnormalizatio ninoverlappingde-scriptorblocksare givesnear-perfectseparationontheoriginal MITpedestriandatabase, soweintroducea more challengingdatasetcontainingover1800anno tatedhumanimageswitha large range a challengingtaskowingtotheirvariableappea ranceandthewiderangeofposesthatthey rstneedisa robustfeaturesetthatallowsthehumanformto bediscriminatedcleanly, eveninclutteredbackgroundsunderdif studytheissueoffeaturesetsforhumandetect ion,showingthatlo-callynormalizedHistogr amofOrientedGradient(HOG)
dient based descriptors, we show experimentally that grids of Histograms of Oriented Gradient (HOG) descriptors sig-nicantly outperform existing feature sets for human detec-tion. We study the inuence of each stage of the computation on performance, concluding that ne-scale gradients, ne orientation binning, relatively coarse spatial binning, and
Domain:
Source:
Link to this page:
Please notify us if you found a problem with this document:
{{id}} {{{paragraph}}}