cvpr18-The INaturalist Species Classification and Detection Dataset

cvpr18-The INaturalist Species Classification and Detection Dataset

ID:40358198

大小:3.10 MB

页数:10页

时间:2019-07-31

cvpr18-The INaturalist Species Classification and Detection Dataset_第1页
cvpr18-The INaturalist Species Classification and Detection Dataset_第2页
cvpr18-The INaturalist Species Classification and Detection Dataset_第3页
cvpr18-The INaturalist Species Classification and Detection Dataset_第4页
cvpr18-The INaturalist Species Classification and Detection Dataset_第5页
资源描述:

《cvpr18-The INaturalist Species Classification and Detection Dataset》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库

1、TheiNaturalistSpeciesClassificationandDetectionDatasetGrantVanHorn1OisinMacAodha1YangSong2YinCui3ChenSun2AlexShepard4HartwigAdam2PietroPerona1SergeBelongie31Caltech2Google3CornellTech4iNaturalistAbstractExistingimageclassificationdatasetsusedincomputervisiontendtohaveauniformdistributiono

2、fimagesacrossobjectcategories.Incontrast,thenaturalworldisheavilyimbalanced,assomespeciesaremoreabundantandeasiertophotographthanothers.ToencouragefurtherprogressinchallengingrealworldconditionswepresenttheiNatural-istspeciesclassificationanddetectiondataset,consistingof859,000imagesfrom

3、over5,000differentspeciesofplantsandanimals.Itfeaturesvisuallysimilarspecies,capturedinawidevarietyofsituations,fromallovertheworld.Im-ageswerecollectedwithdifferentcameratypes,havevary-ingimagequality,featurealargeclassimbalance,andhavebeenverifiedbymultiplecitizenscientists.Wediscussth

4、ecollectionofthedatasetandpresentextensivebaselineex-Figure1.TwovisuallysimilarspeciesfromtheiNat2017dataset.perimentsusingstate-of-the-artcomputervisionclassifica-Throughcloseinspection,wecanseethattheladybugonthelefttionanddetectionmodels.Resultsshowthatcurrentnon-hastwospotswhiletheon

5、eontherighthasseven.ensemblebasedmethodsachieveonly67%toponeclassi-ficationaccuracy,illustratingthedifficultyofthedataset.Itisestimatedthatthenaturalworldcontainsseveralmil-Specifically,weobservepoorresultsforclasseswithsmalllionspecieswitharound1.2millionofthesehavingalreadynumbersoftrain

6、ingexamplessuggestingmoreattentionisbeenformallydescribed[26].Forsomespecies,itmayonlyneededinlow-shotlearning.bepossibletodeterminethespeciesviageneticsorbydis-section.Fortherest,visualidentificationinthewild,whilepossible,canbeextremelychallenging.Thiscanbedueto1.Introductionthesheernu

7、mberofvisuallysimilarcategoriesthatanindi-vidualwouldberequiredtorememberalongwiththechal-Performanceonexistingimageclassificationbench-lenginginter-classsimilarity;seeFig.1.Asaresult,theremarkssuchas[32]isclosetobeingsaturatedbythecur-isacriticalneedforrobustandaccurateautomate

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。