yimingshao1/refracgs_dataset

RefracGS Dataset

Official dataset for RefracGS: Novel View Synthesis Through Refractive Water Surfaces with 3D Gaussian Ray Tracing (ECCV 2026).

Project Page

This is a synthetic benchmark for novel view synthesis through refractive water surfaces. Each scene contains objects placed under water surface, rendered from multiple viewpoints. The dataset is provided in the standard NeRF/Blender format (images + transforms_*.json camera poses).

Dataset Structure

The main dataset (refracgs_synthetic/) contains 3 scenes (desktop, kitchen, toys), each with 30 views:

Directory	Scene	Train	Test
refracgs_synthetic/desktop_30view	desktop	24	6
refracgs_synthetic/kitchen_30view	kitchen	24	6
refracgs_synthetic/toys_30view	toys	24	6

Supplementary data

refracgs_synthetic_supp/ is a supplementary set with sparse-view (9-view) variants of the same 3 scenes, used for the sparse-view experiments in the paper:

Directory	Scene	Train	Test
refracgs_synthetic_supp/desktop_9view	desktop	8	1
refracgs_synthetic_supp/kitchen_9view	kitchen	8	1
refracgs_synthetic_supp/toys_9view	toys	8	1

Each scene directory is organized as:

<scene>/
├── train/                  # training images (PNG, 400x400)
├── test/                   # test images (PNG, 400x400)
├── transforms_train.json   # camera intrinsics + train poses
├── transforms_test.json    # camera intrinsics + test poses
└── transforms_val.json

The transforms_*.json files follow the NeRF/Blender convention: a top-level camera_angle_x field and a frames list with file_path and transform_matrix (camera-to-world, 4x4) per image.

Download

# Clone the whole repository (requires git-lfs)
git lfs install
git clone https://huggingface.co/datasets/yimingshao1/refracgs_dataset

Or with the Hugging Face Hub library:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="yimingshao1/refracgs_dataset",
    repo_type="dataset",
    local_dir="data/refracgs_dataset",
)

Usage with RefracGS

Train RefracGS on this dataset using the official codebase:

# 30-view (dense) scene
CUDA_VISIBLE_DEVICES=0 python train.py --config-name apps/refracgs_synthetic_30view.yaml path=refracgs_dataset/refracgs_synthetic/kitchen_30view out_dir=runs experiment_name=kitchen_30view

# 9-view supplementary (sparse-view) scene
CUDA_VISIBLE_DEVICES=0 python train.py --config-name apps/refracgs_synthetic_9view.yaml path=refracgs_dataset/refracgs_synthetic_supp/kitchen_9view out_dir=runs experiment_name=kitchen_9view

Citation

If you find this dataset useful, please cite our paper:

@inproceedings{shao2026refracgs,
    title     = {RefracGS: Novel View Synthesis Through Refractive Water Surfaces with 3D Gaussian Ray Tracing},
    author    = {Shao, Yiming and Dai, Qiyu and Gao, Chong and Li, Guanbin and Wang, Yequan and Sun, He and Zeng, Qiong and Chen, Baoquan and Chen, Wenzheng},
    booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
    year      = {2026}
}

Acknowledgements

We thank Ashley (@Ashleyyyyy663) for helping us create some of the scenes in this dataset.