CamxTime Evaluation Benchmark

This is the evaluation dataset for the Cam×Time 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.

The dataset contains 32 dynamic scenes, each rendered across a full 120×120 camera×time grid. From this grid, ground-truth videos are extracted for 5 moving-camera evaluation patterns and preprocessed to match the SpaceTimePilot inference format.

---

Folder Structure

CamxTime_eval/
├── full_grid_renders/          Raw full-grid renders (source)
├── eval_input/                 Source input videos + camera files for inference
├── eval_gt/                    Ground-truth pattern videos (native resolution)
├── eval_gt_wan2.1_format/      GT videos preprocessed to match network output
├── process_full_grid_to_gt.py  Script: full_grid_renders → eval_gt
└── preprocess_gt_videos.py     Script: eval_gt → eval_gt_wan2.1_format

full_grid_renders/

Raw renders from a 120×120 camera×time grid per scene.

• 32 scenes, each with 120 camera positions along an arc trajectory
• Per camera: one 120-frame MP4 (1080×1080, 30fps) + camera_data.json with c2w/w2c poses and intrinsics

eval_input/

Source data used as input to the SpaceTimePilot model during inference.

• videos/ — 32 source MP4s (one per scene)
• src_cam/ — per-scene source camera poses (camera_data.json)
• metadata.csv — scene list with text captions

eval_gt/

Ground-truth pattern videos at native resolution (1080×1080, 81 frames), extracted from full_grid_renders by slicing the camera×time grid along 5 trajectories:

Pattern	Camera axis	Time axis
moving_forward	cam 0 → 80	frame 0 → 80
moving_backward	cam 0 → 80	frame 80 → 0
moving_zigzag	cam 0 → 80	0 → 40 → 0
moving_bullettime	cam 0 → 80	frame 40 (frozen)
moving_slowmo	cam 0 → 80	0, 0, 1, 1, …, 40

Generated by process_full_grid_to_gt.py.

eval_gt_wan2.1_format/

GT videos preprocessed to exactly match SpaceTimePilot network output format: 832×480, 81 frames, 30fps H264 (aspect-ratio crop then center-crop from 1080×1080).

Generated by preprocess_gt_videos.py.

---

Generating eval_gt

Script: CamxTime_eval/process_full_grid_to_gt.py

Extracts the 5 GT pattern videos per scene from the full-grid renders. Run from the repo root:

python CamxTime_eval/process_full_grid_to_gt.py \
    --input   CamxTime_eval/full_grid_renders \
    --output  CamxTime_eval/eval_gt \
    --src_cam CamxTime_eval/eval_input/src_cam

Flag	Default	Description
--workers N	ncpu // 8	Parallel scene processes
--threads N	8	ffmpeg threads per scene
--scenes s1 s2	all	Limit to specific scenes

Output per scene: moving_{pattern}.mp4 + .json + .txt + camera_data.json

---

Generating eval_gt_wan2.1_format

Script: CamxTime_eval/preprocess_gt_videos.py

Applies the same spatial transforms as the SpaceTimePilot inference pipeline to the GT videos: scale to cover 832×480 → CenterCrop → pad to 81 frames → 30fps H264.

python CamxTime_eval/preprocess_gt_videos.py \
    --input  CamxTime_eval/eval_gt \
    --output CamxTime_eval/eval_gt_wan2.1_format

Flag	Default	Description
--workers N	min(32, ncpu)	Parallel scene processes
--scenes s1 s2	all	Limit to specific scenes

Both scripts are resumable — already completed scenes are skipped automatically.

---

Citation

@inproceedings{huang2026spacetimopilot,
  title={SpaceTimePilot: Generative Rendering of Dynamic Scenes Across Space and Time},
  author={Huang, Zhening and Jeong, Hyeonho and Chen, Xuelin and Gryaditskaya, Yulia and Wang, Tuanfeng Y. and Lasenby, Joan and Huang, Chun-Hao},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}