Addressing Data Scarcity in Thai Car Recognition Using Few-Shot Learning

This paper aims to address the issues of data scarcity and fine-grained classification in car recognition, specifically in the Thai context, using a few-shot learning model. To achieve this, we conduct experiments to compare several models and establish DN4 as a baseline model. We collect and analyze performance data from various training variables (N-way K-shot) and backbone architectures. Furthermore, we propose the hard task sampler module for few-shot training, which generates more difficult tasks based on the similarity matrix derived from the model's discriminative performance. We test our proposed model on the Comprehensive Cars (CompCars) dataset, which is split into CompCars (non-TH) and CompCars (TH) based on car model popularity in Thailand. By using the best settings from our experiments, our model significantly improves the accuracy of the DN4 baseline.