Data Access
By accessing the data, participants agree to the dataset license and the challenge Terms & Conditions.
Dataset: Native Dynamic Gaussian Sequences
DyGS-Bench focuses on a
native Gaussians-only
setting: the input is a reconstructed dynamic 3D Gaussian sequence (no original multi-view training RGB/video data). Each sequence includes:
Native Gaussian assets
: time-varying Gaussians (per-frame PLY).
Camera set
: calibrated camera parameters for rendering-based evaluation.
Download
The dataset is hosted on HuggingFace:
[https://huggingface.co/datasets/LeeXiangNO1/DyNativeGaussian_sequence](https://huggingface.co/datasets/LeeXiangNO1/DyNativeGaussian_sequence)
Recommended Local Folder Structure (example)
DyNativeGaussian_sequence/
  <scene_name>/
    sparse/
      camera.txt
    gaussians/
      000000.ply
      000001.ply
      ...
        ...
Stage
Stage One
:We will release part of the dataset for training and testing your method. Among the camera viewpoints we provide, two are designated as test viewpoints for platform evaluation. We recommend not attempting to perform supervised training on the test viewpoints, as it is meaningless to boost rankings in the first stage. In the second stage, source code submission will be required, and at that time, the test viewpoints provided by us will not belong to any known subset.
Stage Two
: We will provide additional data sets, as well as notebooks with source code and decoding templates. Please create your own decoding scripts based on these templates. In phase two, you need to provide encoded files and decoding scripts for each data set. We will test the effectiveness of the method on the platform.
Citation
If you use the dataset and benchmark in a publication, please cite the DyGS-Bench paper (citation will be updated here once the preprint is public).