Improving CNN classifiers by estimating test-time priors

TL;DR

This paper introduces a Maximum a Posteriori approach for estimating test-time class priors in CNN classifiers, significantly improving accuracy in fine-grained and unknown prior scenarios.

Contribution

It proposes a novel MAP estimation method with a Dirichlet hyper-prior, enhancing stability and accuracy over existing MLE methods for test-time prior estimation.

Findings

Significant accuracy improvements on fine-grained datasets.

State-of-the-art results on PlantCLEF with unknown priors.

MAP estimation outperforms MLE in accuracy, especially with unknown priors.

Abstract

The problem of different training and test set class priors is addressed in the context of CNN classifiers. We compare two different approaches to estimating the new priors: an existing Maximum Likelihood Estimation approach (optimized by an EM algorithm or by projected gradient descend) and a proposed Maximum a Posteriori approach, which increases the stability of the estimate by introducing a Dirichlet hyper-prior on the class prior probabilities. Experimental results show a significant improvement on the fine-grained classification tasks using known evaluation-time priors, increasing the 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 outperform the competition-winning ensemble of 12 CNNs. The proposed Maximum a Posteriori estimation increases the prediction accuracy by 2.8% on PlantCLEF 2017 and by 1.8% on FGVCx Fungi, where the existing MLE method would lead to a decrease accuracy.