Cycling World Model Dataset

This dataset contains short first-person bicycle riding video clips captured with a camera mounted on the bicycle handlebars, providing the rider's perspective of real-world environments.

It is intended for training and evaluating video world models that predict future frames from past observations. Existing video datasets are often too large for consumer hardware. This dataset offers a lightweight alternative for video world model research.

The clips were extracted from longer recordings of bicycle rides in the same suburban area. While the exact route varies between rides, many locations overlap, allowing models to observe the same environments from slightly different trajectories and under different conditions.

Property	Value
Total duration	2 h 26 min
Clip duration	8 seconds
Clip length	64 frames
Frame rate	8 FPS
Resolution	192 × 108