TJU-DHD

TJU-DHD provides a high-resolution, diverse image dataset for object and pedestrian detection to support development and evaluation of perception models for self-driving vehicles and video surveillance.

Key Features:

• High-Resolution Images: Contains 115,354 images with 52% at 1624×1200 pixels and 48% exceeding 2560×1440 pixels, enabling detection of small objects.
• Diverse Object Representation: Includes 709,330 labeled objects with wide variance in scale and appearance to improve model robustness.
• Environmental Diversity: Captures scenes across different seasons, lighting conditions, and weather scenarios to reflect real-world variability.
• Pedestrian Subset: Provides a specialized subset focused on pedestrian detection across diverse contexts and environments.
• Benchmarking Experiments: Includes experimental validation using RetinaNet, FCOS, FPN, and Cascade R-CNN to assess detection performance.

Scientific Applications:

• Object Detection Research: Facilitates training and evaluation of object detection algorithms using high-resolution and large-scale annotations.
• Pedestrian Detection Benchmarking: Supports development and comparison of pedestrian detectors across varied scales and conditions.
• Autonomous Driving and Surveillance Perception: Enables improvement and validation of perception modules for self-driving vehicles and video surveillance systems.

Methodology:

The dataset was constructed by collecting images that reflect real-world scenarios for self-driving vehicles and video surveillance with emphasis on diversity in object types and environmental conditions; experiments were conducted using RetinaNet, FCOS, FPN, and Cascade R-CNN.