PHOTOREALISTIC AND SYNTHETIC STEREO-DATASET GENERATION METHOD FOR VISUAL ODOMETRY AND DEPTH ESTIMATION

Abstract

Computer vision is a rapidly developing field in modern computer science that deals with various challenging problems. Both mono and stereo imagery data are widely used for tasks such as depth estimation, visual odometry, and SLAM (Simultaneous Localization and Mapping). To ensure the clean verification and robust performance of the resulting software solutions, datasets should contain precise ground truth data. However, creating a real-world stereo dataset is a costly task as it requires stereo cameras and precise hardware for ground truth measurements (such as lidars, lasers, barometers, accelerometers, etc.). The¬se types of hardware are often expensive and not accessible to intermediate users. An alternative approach is to use synthetic datasets, which are collections of computer-gener¬a¬te¬d data designed to mimic real-world data. Synthetic datasets are used to train AI models wh¬e¬n real-world data is not available, or to test the performance of models in simulated environ¬m¬ents. Our method suggests combining real-world data collection with synthetic data gene¬rat¬ion methods to maintain photorealism while gaining the advantages of synthetic data generation flow.