Pedestrian Attribute Recognition Dataset (Pedestrian-1K)

Pedestrian-1K is a specialized synthetic dataset for Pedestrian Attribute Recognition (PAR) and Person Re-Identification (Re-ID) from overhead CCTV perspectives. While existing datasets like PA-100K rely on aging, low-resolution real-world footage with inconsistent labels, Pedestrian-1K provides high-fidelity synthetic subjects from a consistent high-angle security camera geometry. Each image is uniquely annotated with both a clean Natural Language Description (e.g., 'A young adult man with black hair, seen from the side, wearing a navy blue hoodie and black jeans') and a Structured Attribute Map covering gender, age, hair color, clothing type/color, camera angle, and accessories. The dataset features realistic digital noise, scanlines, and compression artifacts to ensure models generalize to real CCTV hardware. Attribute distributions use weighted demographic logic (e.g., elderly subjects lean 80% towards grey hair) for realistic sampling. The annotation format is JSONL with both descriptive strings and categorized attribute labels per image. The open-source sample includes 1,000 images; full packages are available with 10,000 and 50,000 images featuring descriptive annotations and Vision Transformer annotations. Licensed under CC BY 4.0.