Core Concepts
SEVD is a first-of-its-kind multi-view synthetic event-based dataset for autonomous driving and traffic monitoring, offering comprehensive data across diverse lighting, weather, and scene conditions.
Abstract
The SEVD dataset is a significant advancement in the field of synthetic event-based data for autonomous driving and traffic monitoring applications. It provides 27 hours of fixed perception and 31 hours of ego perception event data, complemented by an equal amount of data from other sensor modalities like RGB, depth, optical flow, semantic, and instance segmentation.
The dataset offers a diverse range of recordings featuring various combinations of scenes (urban, suburban, rural, highway), weather (clear, cloudy, wet, rainy, foggy), and lighting conditions (noon, nighttime, twilight). The event data is captured using a strategic multi-camera setup, providing a 360-degree field of view for ego perception and four fixed cameras at key locations like intersections, roundabouts, and underpasses.
SEVD includes extensive annotations, with over 9 million 2D and 3D bounding boxes for six object categories (car, truck, bus, bicycle, motorcycle, and pedestrian). The dataset is segmented into train, validation, and test sets to facilitate model development and evaluation.
The authors establish baseline performance for 2D object detection using state-of-the-art event-based (RED, RVT) and frame-based (YOLOv8) detectors. The results demonstrate the efficacy of the dataset in supporting research on event-based vision for autonomous driving and traffic monitoring tasks. Additionally, the authors conduct experiments to assess the synthetic event-based detector's generalization capabilities on real-world data, providing valuable insights.
Overall, SEVD represents a significant contribution to the field, offering a comprehensive and diverse synthetic event-based dataset that can drive advancements in areas such as object detection, tracking, sensor fusion, and cooperative perception for autonomous systems.
Stats
The dataset includes over 9 million bounding boxes for various traffic participants, including cars, trucks, buses, bicycles, motorcycles, and pedestrians.
Quotes
"SEVD offers raw event streams ⟨x, y, p, t⟩in .npz format alongside their corresponding images."
"SEVD represents a significant advancement as the first-of-its-kind synthetic event-based data providing both ego and fixed perception, featuring a comprehensive range of annotations, extensive recording hours, and diverse driving conditions."