CrowdHuman

Dataset Overview

Dataset Name

CrowdHuman

Dataset Purpose

Used to evaluate detector performance in crowded scenes.

Dataset Scale

• Training set: 15,000 images
• Validation set: 4,370 images
• Test set: 5,000 images
• Total human instances: 470K (training and validation sets)
• Average persons per image: 23

Dataset Characteristics

• Includes various occlusion situations
• Each person instance is annotated with a head box, visible box, and full‑body box

Dataset Structure

Annotation Format

• File format: odgt, each line is a JSON containing all annotations for the corresponding image.

• JSON structure: python JSON{ "ID" : image_filename, "gtboxes" : [gtbox], } gtbox{ "tag" : "person" or "mask", "vbox": [x, y, w, h], "fbox": [x, y, w, h], "hbox": [x, y, w, h], "extra" : extra, "head_attr" : head_attr, } extra{ "ignore": 0 or 1, "box_id": int, "occ": int, } head_attr{ "ignore": 0 or 1, "unsure": int, "occ": int, }

• Annotation notes:

  • tag of mask indicates the box is for crowd/reflective/people‑like objects and should be ignored (extra's ignore = 1)
  • vbox, fbox, hbox represent visible, full, and head boxes respectively

Download Links

• Training set: CrowdHuman_train01.zip
• Training set: CrowdHuman_train02.zip
• Training set: CrowdHuman_train03.zip
• Validation set: CrowdHuman_val.zip
• Training annotations: annotation_train.odgt
• Validation annotations: annotation_val.odgt
• Test set: CrowdHuman_test.zip

Usage Guide

Annotation Conversion

• Use script crowdhuman2coco.py to convert CrowdHuman annotations to COCO format.

Dataset Classes

• Simple PyTorch and MegEngine implementations are provided for reading the CrowdHuman dataset.

• Supported return order: python class CrowdHuman(VisionDataset): supported_order = ( "image", "boxes", "vboxes", "hboxes", "boxes_category", "info", )

• Example usage: python crowdhuman_dataset = CrowdHuman( root=path/to/CrowdHuman, ann_file=path/to/annotations.json, remove_images_without_annotations=True, order=[ image, boxes, boxes_category info ] )