PHDF‑Dataset Overview

Dataset Description

• Type: Pedestrian head detection dataset
• Features: Based on a fisheye camera, bird’s‑eye view
• Scale: Contains 2,201 images, annotated with 13,887 pedestrian heads

Data Collection

• Acquisition: Videos captured with a fisheye camera in three different scenes, supplemented with bird’s‑eye view fisheye videos downloaded from the Bomni‑DB dataset
• Image Processing: Extracted 2,201 frames from video data and resized each image to 512 × 512 pixels
• Annotation: Pedestrian heads in all images are annotated with rectangular boxes

Dataset Composition

• Split: 1,800 images for training, 401 images for testing
• File Structure: Contains train and test folders, each containing images and corresponding XML files that describe head positions

Usage Restrictions

• Purpose: Limited to non‑commercial research