ChenJJ123/d3dgs-benchmark

D3DGS Benchmark: dynamic 3DGS input sequences (N3DV)

Project page · Code · Paper (coming soon)

A dynamic 3D Gaussian splatting scene is a sequence of Gaussian sets, one per frame. How the sets of consecutive frames relate depends on the method that estimated them, and that relation decides what a compression method can exploit. This dataset holds the four input sequence sets of the paper Robust, Estimator-Agnostic Dynamic 3DGS Compression: the six Neural 3D Video scenes, each estimated by four methods, one per correspondence regime. Every codec in the benchmark was run on exactly these files, so new codecs can be compared on identical inputs without training any estimator.

Estimator	Sequence type	Relation between frames	Gaussians per frame	Size
3dgstream	tracked	same Gaussians in the same rows, moved frame to frame (persistent base set)	357 k - 708 k, constant	195 GiB
4dgaussians	tracked	one canonical set deformed to each timestamp	118 k - 132 k, constant	52 GiB
queen	semi-tracked	persistent set with residual updates, additions and removals	248 k - 445 k, varies	132 GiB
3dgs-perframe	untracked	an independent 3DGS optimization per frame, no correspondence	281 k - 845 k, varies	179 GiB

Scenes: coffee_martini, cook_spinach, cut_roasted_beef, flame_salmon, flame_steak, sear_steak, 300 frames each. In total: 24 sequences, 7,200 frames, 558 GiB.

Layout

<estimator>/<scene>/frame_0000.ply ... frame_0299.ply
<estimator>/<scene>/export.json                        # export report (3dgstream, queen)
cameras/<scene>/sparse/0/{cameras,images,points3D}.bin # COLMAP model of the N3DV rig
MANIFEST.csv                                           # every file with its size

Every PLY is binary little-endian with 62 float32 properties per Gaussian: x y z nx ny nz f_dc_0..2 f_rest_0..44 opacity scale_0..2 rot_0..3. Spherical harmonics are degree 3, f_rest is channel-major (15 coefficients of R, then G, then B), and normals are zero. Estimators that store a lower SH degree are zero-padded per channel. This is the layout of the reference 3DGS implementation, so the frames load in common 3DGS viewers and renderers.

Download

One sequence and its cameras:

hf download ChenJJ123/d3dgs-benchmark --repo-type dataset \
    --include "queen/cook_spinach/*" "cameras/cook_spinach/*" --local-dir d3dgs

from huggingface_hub import snapshot_download
path = snapshot_download("ChenJJ123/d3dgs-benchmark", repo_type="dataset",
                         allow_patterns=["queen/cook_spinach/*", "cameras/cook_spinach/*"])

To run the benchmark code, put <estimator>/<scene>/ under data/sequences/ and cameras/<scene>/sparse/0 under data/n3dv/<scene>/colmap/sparse/0 of the repository (see its docs/installation.md).

How the sequences were made

Each estimator was trained with its released code and configuration on the full N3DV rig; the exact commands, commits and patches are in the code repository under estimators/. After export, only these edits were applied:

Edit	Sequences	Extent
SH degree 1 -> 3 zero-padded per channel	3dgstream	layout only
Non-finite values replaced (+-inf -> +-60, NaN -> 0)	3dgstream	< 1e-6 of values
Quaternion sign canonicalized (q and -q are the same rotation)	3dgstream	sign only
Degenerate Gaussians removed (largest log-scale > 2)	queen	about 0.2% of Gaussians
Gaussians with non-finite values removed	queen, coffee_martini	1 Gaussian in each of 127 frames
SH degree 2 -> 3 zero-padded, vertex_id dropped	queen	layout only

3dgstream frames are its persistent base set: the Gaussians of the initial frame as moved by each frame's Neural Transformation Cache, without the Gaussians that frame adds. This keeps a constant count and row correspondence across all 300 frames.

The captured N3DV images are not redistributed. The paper's ground-truth metrics use the held-out camera cam00 of the N3DV release.

License and attribution

The sequences are derived from the Neural 3D Video dataset (Li et al., CVPR 2022, CC BY-NC 4.0) and are released under the same license, for non-commercial use. The estimator implementations keep their own licenses (see THIRD_PARTY_NOTICES.md in the code repository).

Citation

@misc{wang2026robust,
  title  = {Robust, Estimator-Agnostic Dynamic {3DGS} Compression},
  author = {Wang, Chenjunjie and Huang, Zixi and Wang, Yao and Ball{\'e}, Jona},
  year   = {2026}
}
@inproceedings{li2022n3dv,
  title     = {Neural {3D} Video Synthesis from Multi-view Video},
  author    = {Li, Tianye and Slavcheva, Mira and Zollhoefer, Michael and Green, Simon and Lassner, Christoph and Kim, Changil and Schmidt, Tanner and Lovegrove, Steven and Goesele, Michael and Newcombe, Richard and Lv, Zhaoyang},
  booktitle = {CVPR},
  year      = {2022}
}