4DGRT_data / README.md
Shigon's picture
Upload README.md with huggingface_hub
33149a0 verified
|
Raw
History Blame Contribute Delete
1.96 kB
metadata
license: other
task_categories:
  - image-to-video
tags:
  - 4dgrt
  - 4d-gaussian-splatting
  - ray-tracing
  - camera-effects
  - dynamic-scenes
  - synthetic-data
  - novel-view-synthesis

4D-GRT Benchmark Data

This repository contains the synthetic benchmark data for the paper Every Camera Effect, Every Time, All at Once: 4D Gaussian Ray Tracing for Physics-based Camera Effect Data Generation (4D-GRT).

Project Page | arXiv

4D-GRT is a two-stage pipeline that combines 4D Gaussian Splatting with physically-based ray tracing for camera effect simulation. Given multi-view videos, it first reconstructs dynamic scenes, then applies ray tracing to generate videos with controllable, physically accurate camera effects. This repository provides the eight synthetic dynamic indoor scenes constructed as a benchmark to evaluate generated videos across four camera effects: pinhole, fisheye, depth of field, and rolling shutter.

Dataset Description

Each scene is rendered under 4 camera-effect variants (pinhole, fisheye, dof, rolling_shutter), each with 50 cameras and 50 frames at 512x512 resolution, along with a point cloud (points3d_opencv.ply) and transforms_train.json / transforms_test.json camera pose files. A shared effect_params.csv and self_dataset_manifest.json describe the per-scene effect parameters and manifest.

The 8 scenes are: ball, basketball_1, basketball_2, box, cat, cube, lego, plant.

Citation

@misc{liu20254dgrt,
  title={Every Camera Effect, Every Time, All at Once: 4D Gaussian Ray Tracing for Physics-based Camera Effect Data Generation}, 
  author={Yi-Ruei Liu and You-Zhe Xie and Yu-Hsiang Hsu and I-Sheng Fang and Yu-Lun Liu and Jun-Cheng Chen},
  year={2025},
  eprint={2509.10759},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2509.10759}, 
}