Towards Prototype Conformity Loss Functions for Better Outlier Detection in Traffic Sign Image Classification
Abstract: Deep neural networks~(DNNs) generate overconfident outputs even in case of miss-detections caused by abnormal data. Consequently, this can lead to unreliable classifications and, thus, potentially lead to issues in safety-critical applications such as automated driving systems. Recent works propose to detect such anomalous data based on probabilistic methods derived from the DNN's internal activation functions, such as the convolutional neural networks (CNN) backbones. This paper shows that such CNNs cannot semantically disentangle similar classes when trained with conventional cross-entropy loss functions, leading to poor out-of-distribution (OOD) detection while applying probabilistic methods for such a purpose. Therefore, we propose to apply the prototype conformity loss (PCL) function from the literature and show that such a contrastive learning method leads to better OOD detection for traffic sign classification. Furthermore, we propose two novel variations of the PCL, namely weighted PCL (WPCL) and multi-scale PCL (MSPCL), which group similar classes and force the DNN to disentangle them from each other. In contrast to existing contrastive OOD detection literature, we do not rely on complex input transformations or augmentations. We perform our experiments on multiple DNNs and two traffic sign classification datasets, which we test against multiple OOD data sources, such as adversarial and non-adversarial augmentation and real-world OOD data. Based on that, we demonstrate that our PCL variations can achieve superior results in OOD detection when the training dataset includes various similar classes.
Submission Length: Long submission (more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=0qdMfvWXl4
Changes Since Last Submission: Multiple sections of the paper was updated based on the requests and suggestions of the reviewers. The title of the paper was updated to replace "outlier" with "out of distribution" to better address the scope of the paper.
Assigned Action Editor: ~Hongsheng_Li3
Submission Number: 2087
Loading