PROJECT TITLE :

Analysis of Classifier Training on Synthetic Data for Cross-Domain Datasets

ABSTRACT:

Deep learning (DL) presents a number of challenges, one of the most significant of which is the requirement to collect massive amounts of training data. The absence of a dataset that is sufficiently large is one of the most common factors that prevents certain applications from making use of DL. In most cases, the acquisition of the necessary quantities of data requires a significant investment of time, resources, and effort. The use of synthetic images combined with real data is a popular approach that has gained widespread adoption in the scientific community as a means of effectively training a variety of detectors to mitigate the effects of this problem. The potential of training based on synthetic data was investigated by us in this study, which was conducted in the field of intelligent transportation systems. Our primary concentration is on the development of applications for camera-based traffic sign recognition that can be used in advanced driver assistance systems and autonomous vehicles. Structured shadows and gaussian specular highlights are two examples of the innovative augmentation processes that are included in the proposed pipeline for the enhancement of synthetic datasets. The performance of synthetic and real image-based trained models was compared using a well-known deep learning model that was trained using different datasets. In addition to that, a fresh and comprehensive approach to objectively compare these models is suggested here. The description also includes a method known as the semi-supervised errors-guide method, which is used to generate synthetic images. When applied to cross-domain test datasets, our experiments showed that a synthetic image-based approach outperforms in most cases real image-based training (+10% precision for GTSRB dataset). As a result, the generalization of the model is improved, which decreases the cost of acquiring images.