Selection of class-conditional filters for semantic shifted OOD detection

Abstract

Deep neural networks have been deployed in a wide range of applications with remarkable performance but can be easily fooled with data that are out-of-distribution (OOD). Recent works have proposed detection methods for OOD benchmarks consisting of small image datasets from separate domains (i.e. object classification dataset for training samples and digit classification dataset for OOD samples). However, these methods fail for OOD benchmarks consisting of image datasets from the same domain (i.e. Korean food classification dataset for training samples and Italian food classification dataset for OOD samples). To solve this issue, we propose an OOD detection framework that utilizes two simple operations: counting to find class-wise highly activated filters in the last convolution layer, and summation to calculate the confidence score by summation of the activation of highly activated filters. The proposed framework is based on our assumption that given an OOD sample from the same domain, the CNN model will produce similar feature maps through all its filters, while some differences might be found in the feature map from highly activated filters. We show that our method achieves the highest performance for OOD benchmarks consisting of the Food101 dataset, which provides meaningful insights on how this issue, that has been encountered in recent works, can be effectively addressed.