鸟类数据库

The Caltech-UCSD Birds-200-2011 Dataset 用于细粒度检测的鸟类数据库，特别好用，欢迎大家下载使用Acadian FlycatcherCape Glossy Starling吧园American CrowCape May Warbler2N影■圖American GoldfinchCardinal■9■圖圈题American PipitCarolina Wren区圖圖圖圖郾American RedstartCaspian TernCommon RavenHorned GrebeCommon ternHorned larkCommon yellowthroatorned pu斤inCrested aukletHouse Sparrow口剧三Dark eved JuncoHouse wrenFigure 1. CUB-200-2011 Example ImagesPart AttributesPart Attributes PartAttributesBeak HasBillShape,Back Has Backcolor, Breast HasBreastPattern,Has BillcololHasBackPattemHasBreastcolorBelly HasBellyPatternForeHas ForeheadBird(all HasSize, HasShapeHas Bely Colorhead(Right WingRignt EyeThroat Has ThroatcolalHasNapeColor Head HasHeadPatternCrown Has Crown ColorEye HasEyeCololLHasLeg ColorTailHasUpper Tail Color, Wing Has WingPattern, BodHasUnderparts Color,Asunder tail coloHas Wing ColorHasUpperpartscoloHas TailPatternHas WingShapeHasPrimary ColorHasTOILSI(a) Collected Parts(b) Attribute Part AssociationsFigure 2. Collected Parts and Attributes. (a) The 15 part location labels collected for each image. (b) The 28 attribute-groupings thatwere collected for each image, and the associated part for localized allribute detectorsLocate theBEAKClick on the image where the specifieds part is locatedKeyboard shortcuts:left arrow for BACK, right arrow for NEXT.a check this box if more than half ofthe part is not visible in the image1/25(a) Part GUIFigure 3. MTurk GUI for collecting part location labels, deployed on 11, 788 images for 15 different parts and 5 workers per imagePlease draw a rectangle around the bird in theimage. The rectangle should be fit around thebird tightlyGood Rectangles:Bad Rectangles:l Tick any of the boxes below if they are trueabout the image回 Bird is truncated回 Bird is occludeda There is more than one birdd There is no bird in the image14 < Previous一DNe对√ SubmitFigure 4. MTurk GUI for collecting bounding box labels, deployed on 11,788 imagesWhat ls the shape of the blll/beak?1/28Select one. If the beak isn't visible then select "Not visibleConeCurved (up orHookedseabirdNeedleSpatulate Special zedCan't see the brd in the image?aGo Back D Not Visible D Guessing D Probably D DefinitelyClick here to sk p it.ONLY skip if no bird is visible at a‖ In the mace」‖Figure 5. MTurk GUI for collecting attribute labels, deployed on 11, 788 images for 28 different questions and 3 12 binary attributesClass mage Cauntage Size46485052545E5B60051525NITer at ImeneGlSsInaoe sPe IPi331aa)Class Image count(b Image sizesGrappa uncropped Image sze RarlAwerage Part Labaingime35body25而beat33200rght winobreast1002600.10203040506070B0925HarI al ItarU PixE(c) Cropped/Uncropped Image Size ratioAverage part Labeling TimeFigure 6. Dataset Statistics. (a) Distribution of the number of images per class(most classes have 60 images).(b) Distribution of the sizeof each image in pixels (most images are roughly 500X500).(c) Distribution of the ratio of the area of the birds bounding box to the areaof the entire image.(d) The average amount of time it took mturkers to label each partEvaluated on: Predicted Locations, 5 training images/dassEvaluated on: Predicted Locations, max (52)training images/classClassification Accuracy: 6.94%Classification Accuracy: 10.26"%10n12014014016018〔204060801001201401601802040608010012010160180(a) Predicted Locations, 5 Images/Class(b) Predicted Locations, 52 Images/ClassEvalua: ed on: Grourd Truth Locations. 5 training imagesclasEvaluated cn: Ground Truth Locations, max (52 raining images classClassification Acuracy: 10.05%Classification Accur acy: 1731%100。1001201201401601B01802040608010012014016018020406080100120140160180(c) Ground Truth Locations, 5 Images/Class(d) Ground Truth Locations, 52 Images/ClassFigure 7. Categorization Results for 200-way bird species classification. The top 2 images show confusion matrices when using a universalbird detector to detect the most likely location of all parts and then evaluating a multiclass classifier. The bottom 2 images show confusionmatrices when evaluating a multiclass classifier on the ground truth part locations. The 2 images on the left show results with 5 trainingimages per class, and the images on the right show results with 52 training images per classLoss:0.703Loss:0.974Loss:1.093Loss:4.615Loss:5.000Loss:4.231Loss:0.815Loss:1.264Loss:0.987Loss:3.070Loss:4.220Loss:4.003LOSS: 0.928Loss:1.593Los5:0.746Loss:3.836LOSS: 4920Loss:4.615Loss:1,204Loss:1.351Loss:1.338Loss:5.000Loss:3.570Loss:2.992Figure 8 Example Part Detection Results, with good detection results on the left and bad detection results on the right. A loss of 1.0indicates that the predicted part locations are about as good as the average MTurk labeler