The Caltech-UCSD Birds-200-2011 dataset, limited to 30 categories: • Is a significant challenge for bird species classification • Has a fine-grained nature • Presence of occlusions and distractions in the images The semester project, a Kaggle competition of classification : • Produce a model that gives the highest possible accuracy on a test set containing the same categories • Baseline model: A resnet34pretrained on ImageNet achieving a relatively low accuracy of 50%

Assaidbefore,thedatasetprovidedforthecompetitionfacesmultiplechallengestocontrolincluding: • Imagesevenlydistributedamong30birdspeciesclasses • Variouschallenges,includingoff-centerbirdpositioning,occlusionsbyotherobjectssuchastreesorbirdcages,andvariationsinlightingandbackground. • Additionally,thetrainingsetcomprisesover2500imageswithannotations,thevalidationsetcontainsover240imagesformodelevaluation.Thetestset,containingover600images,isusedforfinalmodelinference.

Bird detection and cropping using Deep Learning models: CreationofanewdatasettofacilitatetheclassificationtaskwithobjectdetectionmodelsofbirdswithintheimageswithDetectron2library,leveragingtwoprominentmodelsMaskR-CNNandRetinaNet: • Selectionofthemostconfidentprediction(above85%)ofabirdacrossmodelswiththehighestconfidencedetectionretained • Croppingofalltheimagesinthedataset,therebyfocusingontheregionscontainingbirdsforsubsequentclassification • Preventionofdistortionandensuringuniformityandconsistencybyexpandingtheboundingboxestofitasquareshape • ImplementationofaminimalversionofYOLOv8forbirddetectionduetocomputationalconstraintsandtheprocessingtimesassociatedwithDetectron2 • Augmentationofthetrainingandvalidationdatasetsbycroppingphotoscenteredonthedetectedbirds,therebystreamliningtheoverallworkflowandenhancingtherobustnessofourclassificationmodel

Data augmentation: Toenhancethediversityofthetrainingdataandimprovemodelgeneralization,weapplydataaugmentationtechniquesusingbothtorchvisiontransformsandAlbumentationsincluding: • randomresizingandcropping • horizontalflippingandnormalization • additionalaugmentationssuchasGaussiannoiseandmotionblur.Albumentationslibraryoffersawiderangeofaugmentationoptions,allowingustocreatearobustanddiversetrainingdataset. Dataaugmentationiskeytoavoidoverfittinginourtrainingphase! PS:Weseparatedthedatatransformationsofthetrainingandvalidationsetinordertointroducealeastrestrictingtransformationforthevalidationset.