HUM-CARD: A human crowded annotated real dataset

1. Objective & Motivation
   • The researchers collected and annotated images portraying dense crowds in urban spaces to analyze crowd-related environmental impacts and aid related computer vision tasks.
   • A major goal was to create a dataset that supports researchers developing algorithms for crowd analysis and environmental monitoring.
2. Dataset Composition
   • Includes thousands of real-world urban images, capturing a range of densities and scenes (e.g., streets, transportation hubs).
3. Annotation Process & Quality Control
   • Multiple annotators labeled humans in the scenes to ensure high accuracy.
   • The dataset was refined through consensus mechanisms to improve label consistency.
4. Applications & Benchmarking
   • Beyond a standalone dataset, it is used to benchmark crowd-counting and density estimation models.
   • Demonstrates improvements over existing datasets due to greater complexity and realism.
5. Environmental Insights
   • Analysis explores relationships between crowd density, urban space usage, and environmental stressors, offering a richer context for environmental-management applications.

How Labelbox Was Used

The authors employed Labelbox, a collaborative data‑labeling platform, to efficiently and accurately annotate their dataset:

• They used Labelbox to build the annotation ontology (AO): defining labeling categories (e.g., "main object," sub‑classifications, contextual options) to structure the annotation task)
• The platform enabled sophisticated editor environments, including image segmentation workflows for video frames. This streamlined annotator work and ensured consistent labels across the dataset.
• Key Labelbox features used included:
  • Custom ontologies for complex scene interpretation.
  • Support for bounding boxes, segmentation masks, and multi-label annotations, tailored to dense crowd imagery.
  • Built-in quality control and consensus-based workflows to validate annotator agreement.

The HUM‑CARD dataset provides a rich resource for understanding urban crowding and its environmental dynamics. By leveraging Labelbox’s ontology creation tools, annotation workflows, and quality‑control mechanisms, the authors delivered a high‑quality, densely annotated dataset that supports both environmental studies and crowd‑analysis computer vision research.