CamxTime Evaluation Benchmark

This is the full-grid rendering dataset for the Cam×Time evaluation benchmark introduced in:

SpaceTimePilot: Generative Rendering of Dynamic Scenes Across Space and Time Zhening Huang, Hyeonho Jeong, Xuelin Chen, Yulia Gryaditskaya, Tuanfeng Y. Wang, Joan Lasenby, Chun-Hao Huang CVPR 2026

What is this dataset?

The Cam×Time benchmark evaluates a model's ability to simultaneously control camera viewpoint and temporal motion in a dynamic scene — the core task of SpaceTimePilot.

This dataset contains full-grid renderings of 32 dynamic scenes. Each scene is rendered from a 120×120 camera×time grid (120 camera positions × 120 time steps), which is used to extract ground-truth videos for 5 moving-camera evaluation patterns.

Dataset Structure

• 32 scenes of real-world dynamic content (people, animals, outdoor environments)
• 120 cameras per scene arranged along an arc trajectory
• 120 frames per video (4 seconds @ 30 fps, 1080×1080)
• Camera poses (c2w / w2c matrices) and intrinsics stored in per scene

Evaluation Patterns

From this full grid, 5 ground-truth moved-cam → moved-cam patterns are extracted (81 frames each):

Pattern	Camera	Time
	cam i	frame i
	cam i	frame 80−i
	cam i	zigzag[i] (0→40→0)
	cam i	frame 40 (frozen)
	cam i	slowmo[i] (0,0,1,1,…,40)

How to Use

Extract GT videos

Download a single scene

Citation

License

This dataset is released under the Apache 2.0 License.