VisDrone Dataset

Overview

VisDrone is one of the most comprehensive benchmarks for drone-based computer vision tasks, including object detection and tracking. The dataset was collected by the AISKYEYE team at Lab of Machine Learning and Data Mining, Tianjin University, China.

Dataset Statistics

• Captured by: Various drone-mounted cameras
• Geographic coverage: 14 different cities across China
• Environments: Both urban and country settings
• Object categories: Pedestrians, cars, bicycles, tricycles, and more
• Attributes: Scene visibility, object class, and occlusion information
• Altitude range: 15-180 meters
• Weather conditions: Various weather and lighting scenarios

Tasks Supported

Object Detection in Images (VisDrone-DET)

• Trainset: 1.44 GB
• Valset: 0.07 GB
• Testset-dev: 0.28 GB (Ground truth available)
• Testset-challenge: 0.28 GB

Object Detection in Videos (VisDrone-VID)

• Trainset: 7.53 GB
• Valset: 1.49 GB
• Testset-dev: 2.14 GB (Ground truth available)
• Testset-challenge: 2.70 GB

Single-Object Tracking (VisDrone-SOT)

• Trainset: ~20.37 GB (split into two parts)
• Valset: 1.29 GB
• Testset-dev: 11.27 GB (Ground truth available)
• Testset-challenge: ~34.71 GB (split into two parts)

Multi-Object Tracking (VisDrone-MOT)

• Trainset: 7.53 GB
• Valset: 1.48 GB
• Testset-dev: 2.14 GB (Ground truth available)
• Testset-challenge: 2.70 GB

Strengths and Limitations

Characteristics for Drone Deployment

• Collected using various drone platforms (different models)
• Captured under various weather and lighting conditions
• Includes sparse and crowded scenes
• Provides annotations for common objects of interest in aerial surveillance