Learning-Deep-Learning

WiderPerson: A Diverse Dataset for Dense Pedestrian Detection in the Wild

July 2020

tl;dr: A relatively scale (8k training images) dataset for crowded/dense human detection.

Overall impression

Overall not quite impressive. It fails to cite a closely related dataset CrowdHuman, and ablation study of the issue is not as extensive as well.

Key ideas

• 30 persons per image.
• Annotate top of the head and middle of the feet (similar to CityPerson). The bbox is automatically generated with aspect ratio of 0.41. This is
• Difficulty: > 100 pixel (easy), > 50 pixel (medium), > 20 pixel (hard). Similar to WiderFace.
• NMS is a problem in crowded scenes, but it is not handled in this paper. Maybe try Visibility Guided NMS.

Technical details

• Use pHash to avoid duplication of images.
• Annotation tool with examples in the GUI.
  • 
• Evaluation metric: MR

Notes

• Tsinghua-Daimler datasets for cyclists
  • Bounding Box based labels are provided for the classes: (“pedestrian”, “cyclist”, “motorcyclist”, “tricyclist”, “wheelchairuser”, “mopedrider”).
• The EuroCity Persons Dataset: A Novel Benchmark for Object Detection T-PAMI 2019