Sphar: 7,759 Surveillance-Perspective Videos for Human Action Recognition

Sphar contains 7,759 spatio-temporally cropped videos categorized into 14 human action classes, curated by AlexanderMelde and updated in April 2025. The collection aggregates footage from multiple sources specifically filtered to simulate surveillance-camera viewpoints.

Use Cases

• Classifying human behaviors across 14 categories in CCTV-style footage
• Training action recognition models using spatio-temporally cropped video inputs
• Benchmarking video prediction algorithms on surveillance-perspective data

Strengths

• 7,759 video records
• 14 distinct action categories
• Spatio-temporal cropping for action isolation
• Consistent surveillance-specific camera angles

Limitations

• Aggregated from multiple sources which may result in inconsistent lighting and resolution
• Unknown license status
• Lack of detailed metadata regarding the original source environments

Provenance

Source

AlexanderMelde via GitHub

Collection Method

Aggregated from multiple sources and manually cropped spatio-temporally

Freshness

Last updated April 2025.