Improving CNN Classifiers by Estimating Test-Time Priors

The problem of different training and test set class priors is addressed in the context of CNN classifiers. We compare two approaches to the estimation of the unknown test priors: an existing Maximum Likelihood Estimation (MLE) method and a proposed Maximum a Posteriori (MAP) approach introducing a Dirichlet hyper-prior on the class prior probabilities. Experimental results show a significant improvement in the fine-grained classification tasks using known evaluation-time priors, increasing top-1 accuracy by 4.0% on the FGVC iNaturalist 2018 validation set and by 3.9% on the FGVCx Fungi 2018 validation set. Estimation of the unknown test set priors noticeably increases the accuracy on the PlantCLEF dataset, allowing a single CNN model to achieve state-of-the-art results and to outperform the competition-winning ensemble of 12 CNNs. The proposed MAP estimation increases the prediction accuracy by 2.8% on PlantCLEF 2017 and by 1.8% on FGVCx Fungi, where the MLE method decreases accuracy.