Geometric loss functions for camera pose regression论文的报告ppt，纯个人制作，原创。on: IntroductionWhat is this article talking about- pose regressionGiving a single imageOutput where and which pose people usedcould take those picturesApplications■ Autonomous vehiclesUnmanned aerial vehiclesAugmented realityEtcon Introductiontraditional methodimage retrievaldescriptor matchingclassification networksWeaknessLimitation of storage space and Computing ability■ RobustnessLow precision: on: IntroductionThis paperEnd-to-end deep learningNot require memory linearly proportional to thesize of the sceneImprove the precisionRemove the hyperparameterscapable of being applied to any neural networktrained through back propagationImprove the performance of PoseNet withgeometrically formed loss function09model02. ModelPose RepresentationModelArchitectureLoSS Function709 ModelTips for deep architecture forArchitectureimage classification modelGoogleRemove the final linearLeNetregression and soft max laversused for classificationAppend a linear regressionlayer which designed toConv1×1+1(S)output a seven dimensional≡pose vector representing≡position and orientationMaxPool3x3+1(S)Insert a normalization layer tonormalize the fourdimensional quaternionSoftmaxActivationorientation vector to unitlength809ModelPose RepresentationQuaternions- using 4 components to describe the rotationQuaternions are favourable because arbitrary for dimensional values areeasily mapped to Legitimate rotations by normalizing them to unit length.. Simpler process than the orthonormalization required of rotation matricesqUaternions are a continuous and smooth representation of rotation.q=s+ai +yj+zk, (S, 3,ER)2=k2=ik=-1ModelLoss functionLearningLearningLearning anfromPosition andSimultaneouslyOptimalGeometricRegressionorientationweightingReprojectionnormerrorConstraion all Design a loss Formulate itConvertsUsing L1functionusIngrotationquaternions toonewhich is able homoscedasticandnormhemisphere to learn both uncertaintytranslationsuch thatposition and which we canquantitiesthere is aorientationlearn usingInto imageunique valueprobabilisticcoordinatesfor eachdeep learningrotation10