VEHICLE DETECTION AND IDENTIFICATION WITH SMALL DATASET USING FEW-SHOT LEARNING

Abstract

Vehicle detection and identification serve an important role in employing autonomous vehicle classification. However, most deep learning methods for vehicle detection rely on large number of datasets for the training to perform well. The dataset shortage has become a main problem to acquiring a high accuracy detection. A detection method that could adapt with small samples could be a powerful solution for this problem. One of the popular methods is the Few-Shot Learning (FSL) algorithm. This method utilizes meta learning that can deliver more accurate detection with a small number of datasets. This research aimed to assess the effectiveness of the FSL algorithm in classifying road vehicles in Indonesia. Road vehicles categorization which based on the Decree of the Indonesian Minister of Public Works is unique, whereby it differentiates the trucks based on their axles. The datasets for these trucks are rarely available. This research work has employed a small amount of data to the FSL algorithm with reweighting module (FSRW) for classification. The results have demonstrated that FSRW method with fine tuning achieved better results in road vehicle identification with around 33% increase compared to the baseline YOLO method in categorizing road vehicles in Indonesia context.

Keywords

• few-shot object detection, few-shot learning, vehicle identification, vehicle classification.