Formanycomputervisionproblems,themosttimeconsumingcomponentconsistsofnearestneighbormatchinginhigh-dimensionalspaces.Therearenoknownexactalgorithmsforsolvingthesehigh-dimensionalproblemsthatarefasterthanlinearsearch.Approximatealgorithmsareknowntoprovidelargespeedupswithonlyminorlossinaccuracy,butmanysuchalgorithmshavebeenpublishedwithonlyminimalguidanceonselectinganalgorithmanditsparametersforanygivenproblem.Inthispaper,wedescribeasystemthatanswersthequestion,“Whatisthefastestapproximatenearest-neighboralgorithmformydata?”Oursystemwilltakeanygivendatasetanddesireddegreeofprecisionandusethesetoautomaticallydeterminethebestalgorithmandparametervalues.Wealsodescribeanewalgorithmthatappliesprioritysearchonhierarchicalk-meanstrees,whichwehavefoundtoprovidethebestknownperformanceonmanydatasets.Aftertestingarangeofalternatives,wehavefoundthatmultiplerandomizedk-dtreesprovidethebestperformanceforotherdatasets.Wearereleasingpublicdomaincodethatimplementstheseapproaches.Thislibraryprovidesaboutoneorderofmagnitudeimprovementinquerytimeoverthebestpreviouslyavailablesoftwareandprovidesfullyautomatedparameterselection.
1