Trainingagenericobjectnessmeasuretoproduceasmallsetofcandidateobjectwindows,hasbeenshowntospeeduptheclassicalslidingwindowobjectdetectionparadigm.Weobservethatgenericobjectswithwell-definedclosedboundarycanbediscriminatedbylookingatthenormofgradients,withasuitableresizingoftheircorrespondingimagewindowsintoasmallfixedsize.Basedonthisobservationandcomputationalreasons,weproposetoresizethewindowto8×8andusethenormofthegradientsasasimple64Dfeaturetodescribeit,forexplicitlytrainingagenericobjectnessmeasure.Wefurthershowhowthebinarizedversionofthisfeature,namelybinarizednormedgradients(BING),canbeusedforefficientobjectnessestimation,whichrequiresonlyafewatomicoperations(e.g.ADD,BITWISESHIFT,etc.).ExperimentsonthechallengingPASCALVOC2007datasetshowthatourmethodefficiently(300fpsonasinglelaptopCPU)generatesasmallsetofcategory-independent,highqualityobjectwindows,yielding96.2%objectdetectionrate(DR)with1,000proposals.IncreasingthenumbersofproposalsandcolorspacesforcomputingBINGfeatures,ourperformancecanbefurtherimprovedto99.5%DR
1