VSAI Dataset

Dataset Composition

The VSAI dataset contains an impressive 49,712 vehicle instances meticulously annotated using oriented bounding boxes and arbitrary quadrilateral bounding boxes. This level of precision in annotation accommodates varying object shapes and positions in aerial views, enhancing the dataset’s utility for training advanced object detection algorithms. It includes annotations for:

• Small Vehicles: 47,519 instances
• Large Vehicles: 2,193 instances

To enhance model robustness, the dataset also includes detailed annotations of occlusion rates, helping algorithms generalize better under challenging scenarios like partial visibility caused by environmental factors or other objects.

Key Features

• Diverse Capture Conditions: The images in VSAI are captured across multiple drone platforms, which vary in altitude, camera angles, and flight paths.
• Complex Scenarios: The dataset spans a wide range of environments, including urban centers, rural landscapes, industrial areas, and highways, representing real-world complexities.
• Varied Temporal and Weather Conditions: Images have been captured during different times of the day and under diverse weather conditions, including clear skies, overcast, rainy, and foggy environments.
• High Annotation Accuracy: Vehicle instances are annotated with highly precise bounding boxes, including both oriented and arbitrary quadrilateral shapes, to reflect the true contours of vehicles as seen from an aerial perspective.
• Occlusion Annotation: Objects in the dataset are also labeled with their occlusion levels, providing additional challenges for model training and improving detection performance in scenarios with partially visible vehicles.