TJU-DHD: A Diverse High-Resolution Dataset for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Object detection is crucial for autonomous vehicles and surveillance, but current methods struggle with small objects.
• Existing datasets lack the scale, diversity, and resolution needed for robust real-world performance.
• Public datasets often do not focus on specific, challenging scenarios.

Purpose of the Study:

• To introduce the TJU-DHD dataset, a large-scale, high-resolution resource for object and pedestrian detection.
• To address the limitations of existing datasets in terms of image resolution, object diversity, and environmental variability.
• To facilitate advancements in perception systems for autonomous driving and video surveillance.

Main Methods:

• Construction of the TJU-DHD dataset with over 115,000 high-resolution images (1624x1200 and 2560x1440+ pixels).
• Annotation of 709,330 objects with significant variations in scale and appearance.
• Inclusion of diverse seasonal, illumination, and weather conditions.
• Development of a supplementary diverse pedestrian dataset.
• Evaluation of four object detectors (RetinaNet, FCOS, FPN, Cascade R-CNN) on the new dataset.

Main Results:

• The TJU-DHD dataset provides unprecedented scale and resolution for object detection research.
• Experiments demonstrate the dataset's utility in evaluating various object detection algorithms.
• The dataset's diversity challenges current models, highlighting areas for improvement.

Conclusions:

• The TJU-DHD dataset is a valuable resource for advancing object and pedestrian detection research.
• High-resolution, diverse datasets are essential for developing practical perception systems.
• The dataset is expected to spur innovation in autonomous driving and surveillance technologies.