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

Abstract

A major challenges of deep learning (DL) is the necessity to collect huge amounts of training data. Often, the lack of a sufficiently large dataset discourages the use of DL in certain applications. Typically, acquiring the required amounts of data costs considerable time, material and effort. To mitigate this problem, the use of synthetic images combined with real data is a popular approach, widely adopted in the scientific community to effectively train various detectors. In this study, we examined the potential of synthetic data-based training in the field of intelligent transportation systems. Our focus is on camera-based traffic sign recognition applications for advanced driver assistance systems and autonomous driving. The proposed augmentation pipeline of synthetic datasets includes novel augmentation processes such as structured shadows and gaussian specular highlights. A wellknown DL model was trained with different datasets to compare the performance of synthetic and real image-based trained models. Additionally, a new, detailed method to objectively compare these models is proposed. Synthetic images are generated using a semi-supervised errors-guide method which is also described. Our experiments showed that a synthetic image-based approach outperforms in most cases real image-based training when applied to cross-domain test datasets (+10% precision for GTSRB dataset) and consequently, the generalization of the model is improved decreasing the cost of acquiring images.

Leveraging Synthetic Data for DNN-Based Visual Analysis of Passenger Seats

“…Some other works apply image augmentations to the generated samples instead of focusing on the 3D scene generation and rendering step. [32] generate new augmented synthetic samples using a semisupervised errors-guide method to improve the DNN accuracy on cross-domain datasets.…”

A Learnable Discrete-Prior Fusion Autoencoder with Contrastive Learning for Tabular Data Synthesis

“…Synthetic data has demonstrated robust results in overcoming data limitations for various tasks such as dataset balancing (Xiao, Wu, and Lin 2021), data analysis (Cortés et al 2022;Zhang et al 2019;Lou et al 2022) and privacy preservation (Faisal et al 2022;Liu et al 2022). Nowadays, the main popular methods to synthesize tabular data by using deep learning are based on GANs.…”

“…Nevertheless, owing to the ever-present irreconcilable conflict between data leakage risk and availability, such approaches still have a loss of raw information, which results in poorly trained AI models (Cheng et al 2022). The adoption of synthetic data for machine learning has gained significant traction in recent years (Cortés et al 2022). Researchers have been exploring a well-secured approach to generate synthetic data that closely resembles real data, often referred to as "almost-butnot-quite replica data" (Ganev, Oprisanu, and De Cristofaro 2022).…”

Methodology for generating synthetic labeled datasets for visual container inspection

“…Moreover, comparing metrics between other algorithms is next to impossible. Synthetic data generation has proven to be a good approach to tackle the lack of data (Cortes et al, 2022). Data Augmentation is the first step that most of the current algorithms use in order to increment the amount of data with few effort (Zoph et al, 2020;Shin et al, 2020).…”