EVolSplat4D Example Scenes

Prepared inference examples for EVolSplat4D: Efficient Volume-based Gaussian Splatting for 4D Urban Scene Synthesis (IJCV 2026).

Code · Paper · Pretrained checkpoint

This repository contains 16 processed scenes and 900 RGB frames: 5 dynamic Waymo scenes, 5 dynamic PandaSet scenes, and 6 static Waymo scenes. Each scene is supplied as a separate .tar.gz archive. Download and extract the archives, then use the EVolSplat4D data parser. These archives are not a Hugging Face datasets.load_dataset() table or a WebDataset collection; the dataset viewer is disabled.

Contents

README.md
LICENSE
SHA256SUMS
dynamic/
  scene_*.tar.gz          # 5 Waymo scenes, Drop80
  pandaset_*.tar.gz       # 5 PandaSet scenes, Drop80
static/
  scene_*.tar.gz          # 6 Waymo scenes, Drop50

Each archive extracts to a directory named after its scene:

<scene>/
  transforms.json
  front_images/                  # RGB images
  feature_map/                   # Precomputed DINO features (.pt)
  input_pcd/
    Drop80/Static.npz            # Dynamic scene point-cloud prior
    Drop50/Static.npz            # Static scene point-cloud prior
    track_info.pth               # Dynamic object tracks, dynamic scenes only

The point-cloud prior contains points and features arrays. Camera intrinsics, poses, image paths, scene bounds, and prior paths are stored in transforms.json. Dynamic archives retain only the inputs needed by the default Drop80 inference path; they do not include Drop50 priors or preprocessing intermediates. Static archives retain the existing processed files, including auxiliary depth, semantic masks, sky masks, videos, and any other point-cloud priors. Their metadata omits optional references to files that do not exist. The image, feature, point-cloud, and trajectory values are unchanged.

Download and extract

Install the Hugging Face CLI with pip install -U huggingface_hub. This repository is public; you can download the example scenes without logging in to Hugging Face.

Download all scenes and accompanying documentation:

hf download cookiemiao/EVolSplat4D --repo-type dataset --local-dir example_data
(cd example_data && sha256sum -c SHA256SUMS)
for archive in example_data/dynamic/*.tar.gz example_data/static/*.tar.gz; do
  tar -xzf "$archive" -C "$(dirname "$archive")"
done

Download only dynamic scenes (replace dynamic/* with static/* for static scenes):

hf download cookiemiao/EVolSplat4D --repo-type dataset \
  --include "dynamic/*" --local-dir example_data
hf download cookiemiao/EVolSplat4D README.md LICENSE SHA256SUMS \
  --repo-type dataset --local-dir example_data
(cd example_data && sha256sum -c --ignore-missing SHA256SUMS)

Download and extract one scene:

hf download cookiemiao/EVolSplat4D dynamic/scene_121_000_060.tar.gz \
  README.md LICENSE SHA256SUMS --repo-type dataset --local-dir example_data
(cd example_data && sha256sum -c --ignore-missing SHA256SUMS)
tar -xzf example_data/dynamic/scene_121_000_060.tar.gz -C example_data/dynamic

Run commands below from the EVolSplat4D code repository, with example_data/ downloaded into that directory. Follow the code repository's installation instructions first.

Inference

Download the shared 40,000-step checkpoint:

hf download cookiemiao/EVolSplat4D pretrain_waymo.ckpt \
  --repo-type model --local-dir weight

Dynamic Waymo or PandaSet, using Drop80:

python nerfstudio/scripts/infer_4d.py evolsplat4d \
  --load_checkpoint weight/pretrain_waymo.ckpt \
  --config_file config/evolsplat4d_dynamic.yaml \
  --pipeline.model.freeze_volume=True \
  evolsplat4d-zeroshot-data \
  --data example_data/dynamic/scene_121_000_060 \
  --eval_mode drop80

For PandaSet, replace the scene path with an extracted pandaset_* directory. For static Waymo, use the following command after extracting a static archive:

python nerfstudio/scripts/infer_4d.py evolsplat4d \
  --load_checkpoint weight/pretrain_waymo.ckpt \
  --config_file config/evolsplat4d_static.yaml \
  --pipeline.model.freeze_volume=True \
  evolsplat4d-zeroshot-data \
  --data example_data/static/scene_003_039_089 \
  --eval_mode drop50

Static inference disables dynamic modeling and does not require trajectories. Both modes use the same checkpoint. Outputs are written to Zeroshot/extrap/<scene>/; running the same scene again replaces that scene's existing output directory.

Evaluation protocol

Frame indices below are zero-based and follow the parser's sorted image order.

Group	Scenes	Frames per scene	Input views	Evaluation views
Dynamic Waymo, Drop80	5	60	13	46
Dynamic PandaSet, Drop80	5	60	13	46
Static Waymo, Drop50	6	50	26	23

Dynamic Drop80 uses indices 0, 5, …, 55 and 59 as inputs, evaluates the remaining indices except 58, and preserves all 60 images and feature files for parser compatibility. Static Drop50 uses even indices plus 49 as inputs, evaluates odd indices 1–45, and excludes 47. These prepared examples contain 598 evaluation views in total. They are inference examples, not a training split; no training scenes, KITTI, or KITTI-360 data are included.

Data sources and terms

Waymo scenes are derived from the Waymo Open Dataset, provided by Waymo LLC under the Waymo Dataset License Agreement for Non-Commercial Use. Access and use of these processed scenes are governed by that agreement. Redistribution is limited to recipients who have registered with Waymo and accepted its terms.

PandaSet scenes are derived from PandaSet, provided by Scale AI, Inc. and Hesai Photonics Technology Co., Ltd. PandaSet is provided under CC BY 4.0 with the additional Dataset Terms supplied with the source data. Both source terms are included in LICENSE. EVolSplat4D preprocessing prepares images, DINO features, point-cloud priors, camera metadata, and object tracks; it does not replace the upstream data licenses with the code license.

These are selected driving scenes, not a representative sample of all environments, traffic, weather, or sensor configurations. The examples support research reproduction and are not a safety validation dataset.

Citation

@article{miao2026evolsplat4d,
  title={{EVolSplat4D}: Efficient Volume-based Gaussian Splatting for {4D} Urban Scene Synthesis},
  author={Miao, Sheng and Li, Sijin and Wang, Pan and Bai, Dongfeng and Liu, Bingbing and Wang, Yue and Geiger, Andreas and Liao, Yiyi},
  journal={International Journal of Computer Vision},
  year={2026},
  note={Accepted for publication},
  url={https://arxiv.org/abs/2601.15951}
}

Please also acknowledge the original Waymo Open Dataset and PandaSet when using their scenes.

Scene inventory

The table below lists all downloadable archives. SHA-256 checksums are provided in SHA256SUMS.

Archive	Dataset	Frames	Size (GiB)
dynamic/scene_121_000_060.tar.gz	Waymo	60	0.518
dynamic/scene_130_130_190.tar.gz	Waymo	60	0.519
dynamic/scene_140_040_100.tar.gz	Waymo	60	0.518
dynamic/scene_146_020_080.tar.gz	Waymo	60	0.517
dynamic/scene_146_090_150.tar.gz	Waymo	60	0.517
dynamic/pandaset_001_018_078.tar.gz	PandaSet	60	0.455
dynamic/pandaset_016_000_060.tar.gz	PandaSet	60	0.451
dynamic/pandaset_037_018_078.tar.gz	PandaSet	60	0.448
dynamic/pandaset_101_000_060.tar.gz	PandaSet	60	0.455
dynamic/pandaset_110_020_080.tar.gz	PandaSet	60	0.440
static/scene_003_039_089.tar.gz	Waymo	50	1.969
static/scene_226_031_081.tar.gz	Waymo	50	2.015
static/scene_245_098_148.tar.gz	Waymo	50	1.997
static/scene_246_050_100.tar.gz	Waymo	50	2.004
static/scene_271_110_160.tar.gz	Waymo	50	1.955
static/scene_297_070_120.tar.gz	Waymo	50	1.992

Total compressed size: 16.77 GiB.