Few-Shot Learning for Vehicle Make Model Recognition: Weight Imprinting vs. Nearest Class Mean Classifiers

摘要

In this paper, we study few-shot learning problem with an application to vehicle make and model recognition task in traffic surveillance images. While they may be considered as a robust alternative to ALPR based vehicle model recognition, image based vehicle recognition systems typically fails when an unseen vehicle appears in the scene. Few-shot learning methods may offer a promising solution in resolving this unseen class problem for vehicle model recognition task. In this study, we compare two popular few-shot learning approaches on vehicle model recognition task, namely weight imprinting and nearest class mean classifier. Objective of the proposed approach is to yield good classification performance on the novel classes while keeping the high accuracy rate of the base (existing) classes. We evaluate the effectiveness of the proposed approach using 8423 test images belonging to 100 novel and 100 existing categories. Experimental results have shown that the nearest class mean classifier outperforms weight imprinting on this task with an overall accuracy rate of 82% compared to 65% of weight imprinting.