Recognizing Traffic Signs using Fine-tuning Based Few-Shot Object Detection

Abstract

The most critical task for the Advanced Driver Assistance System (ADAS) which is generally used in autonomous vehicles is to develop a reliable and fast Traffic Sign Recognition (TSR) system. TSR identifies the traffic sign from an image and then determines its category. The majority of widely used TSR techniques that rely on deep convolutional neural networks (DCNNs) emphasize on discriminating feature learning against visual differences of different traffic signs. But these techniques perform poorly if the number of samples available for each of the category is limited to model training. To overcome this problem, few-shot learning can be used where the approach focuses on learning common but distinctive qualities of class-specific objects with few training samples, as opposed to depending heavily on supervision to learn discriminating features. In this work, we have used fine-tuning approach for few-shot learning in order to recognize traffic signs with only a limited number of samples per category. We have introduced Domain Adaptation, Warm Model, Pseudo-Support Set and Instance-Level Feature Normalization in our base architec ture. Our model outperformed all state-of-the-art (SOTA) architectures for few-shot learning across different shot settings, including 2, 3, 5, and 10 shots. Particularly, our model achieved remarkable results in 3-shot and 5-shot scenarios, with an addi tional mAP improvement of 3.53 and 3.73, respectivel