DL3DV-GS-960P Dataset

DL3DV-GS-960P dataset contains 6939 samples of undistorted images, camera poses, and pre-trained 3DGS, under 960P resolution. This dataset is originated from DL3DV, and post processed by FCGS. More information can be found in the Appendix of FCGS paper.

Download

If you have enough space, you can use git to download a dataset from huggingface.

Or you can use downloading script that provides more flexibility.

0. git clone https://github.com/YihangChen-ee/FCGS.git. Downloading scripts are put under dataset folder in this repo.
1. Python script dataset/download_DL3DV-GS-960P.py to download the dataset.
2. dataset/hash_name_train.txt containing hash ids for training scenes. -- Split by FCGS
3. dataset/hash_name_test.txt containing hash ids for testing scenes. -- Split by FCGS

The usage of the download script:

usage: download_DL3DV-GS-960P.py [-h] --odir ODIR --subset {/,1K,2K,3K,4K,5K,6K,7K} --file_type {3DGS,imgs_undist} [--hash HASH] --split {train,test}

optional arguments:
  -h, --help            show this help message and exit
  --odir ODIR           output directory
  --subset {/,1K,2K,3K,4K,5K,6K,7K}
                        The subset of the benchmark to download. A / by default indicates downloading all subsets one by one. 
  --file_type {3DGS,imgs_undist}
                        The file type to download. 
                        1) 3DGS: pre-trained 3DGS. 
                        2) imgs_undist: both undistorted images and camera poses
  --hash                If set hash, this is the hash code of the scene to download
  --split {train,test}   
                        The training or testing split. For the testing split, chkpnt30000.pth is also provided in the 3DGS file_type.

Here are some examples:

# Make sure you have applied for the access.

# Download pretrained 3DGS, 1K subset, output to ./DL3DV-GS-960P directory, for training
python download_DL3DV-GS-960P.py --odir ./DL3DV-GS-960P --subset 1K --file_type 3DGS --split train

# Download both undistorted images and camera poses, 6K subset, output to ./DL3DV-GS-960P directory, for testing
python download_DL3DV-GS-960P.py --odir ./DL3DV-GS-960P --subset 6K --file_type imgs_undist --split test

You can also download a specific scene with its hash. The scene-hash pair visualization can be found here.

python download_DL3DV-GS-960P.py --odir ./DL3DV-GS-960P --hash a38bb446a5e6117ca5cc44fb0809a37ac59a8cfb7093be6d0bc5a5b32aee156e --file_type imgs_undist --split train 

BibTeX

If you found DL3DV-GS-960P Dataset useful, please cite:.

@inproceedings{fcgs2025,
  title={Fast Feedforward 3D Gaussian Splatting Compression},
  author={Chen, Yihang and Wu, Qianyi and Li, Mengyao and Lin, Weiyao and Harandi, Mehrtash and Cai, Jianfei},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025}
}

@inproceedings{ling2024dl3dv,
  title={Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision},
  author={Ling, Lu and Sheng, Yichen and Tu, Zhi and Zhao, Wentian and Xin, Cheng and Wan, Kun and Yu, Lantao and Guo, Qianyu and Yu, Zixun and Lu, Yawen and others},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={22160--22169},
  year={2024}
}