Datasets:

GaussianWorld
/

scene_splat_7k

license: cc-by-sa-4.0
task_categories:
  - other
language:
  - en
tags:
  - indoor
  - 3dgs
  - pretraining
size_categories:
  - 1K<n<10K
license_name: scene-splat-7k-license
pretty_name: SceneSplat-7K
configs:
  - config_name: ARKitScenesGS
    data_files:
      - split: train
        path: statistics/arkitscenes_3dgs_runs.csv
  - config_name: HyperSimGS
    data_files:
      - split: train
        path: statistics/hypersim_mcmc_3dgs_runs.csv
  - config_name: Matterport3D_Region_GS
    data_files:
      - split: train
        path: statistics/matterport3d_region_mcmc_3dgs_runs.csv
  - config_name: Matterport3D_Scene_GS
    data_files:
      - split: train
        path: statistics/matterport3d_scene_mcmc_3dgs_runs.csv
  - config_name: ReplicaGS
    data_files:
      - split: train
        path: statistics/replica_mcmc_3dgs_runs.csv
  - config_name: ScanNetPPGS-V1
    data_files:
      - split: train
        path: statistics/scannetpp_v1_3dgs_runs.csv
  - config_name: ScanNetPPGS-V2
    data_files:
      - split: train
        path: statistics/scannetpp_v2_3dgs_runs.csv
  - config_name: ScanNetGS
    data_files:
      - split: train
        path: statistics/scannet_3dgs_runs.csv
  - config_name: 3RScanGS
    data_files:
      - split: train
        path: statistics/3rscan_mcmc_3dgs_runs.csv
extra_gated_prompt: >
  SceneSplat-7K is built upon multiple existing 3D datasets, each with their own
  licensing requirements. We've carefully structured our distribution approach
  to respect all original licenses while making our dataset accessible to the
  research community.


  Before we are able to offer you access to the SceneSplat-7K dataset, please
  agree that you will use the dataset in accordance with the original
  licenses/terms.


  If you find our work helpful, please consider citing our paper.


  @inproceedings{li2025scenesplat,
    title={SceneSplat: Gaussian Splatting-based Scene Understanding With Vision-Language Pretraining},
    author={Li, Yue and Ma, Qi and Yang, Runyi and Li, Huapeng and Ma, Mengjiao and Ren, Bin and Popovic, Nikola and Sebe, Nicu and Konukoglu, Ender and Gevers, Theo and others},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year={2025}
  }
extra_gated_fields:
  Full name: text
  Current affiliation: text
  Type of Affiliation:
    type: select
    options:
      - Academia
      - Industry
      - label: Other
        value: other
  Institutional email: text
  Please explain your intended research use: text
  I agree to all the terms outlined by original datasets: checkbox
  I agree to use this the data for non-commercial, academic purposes only: checkbox

Updates

The HyperSimGS and ReplicaGS data have been updated with improved quality.
The ScanNetGS and ScanNetPPGS-V2 data have been updated with improved quality.
The ARKitScenes 3DGS data have been updated with improved quality.
The Matterport3D 3DGS data have been updated with improved quality. The new GS files were optimized from each complete scene rather than from a region. The preprocessed npy version is available at Matterport3D 3DGS. Note that the test_eval folder contains 3DGS data corresponding to the test split of the region_segmentations version, on which we reported results.

SceneSplat-7K

We propose the SceneSplat-7K dataset, which includes indoor 3D Gaussian Splatting scenes optimized from ScanNet, ScanNet++, ScanNet++ v2, Replica, Hypersim, 3RScan, ARKitScenes, and Matterport3D. In total, the dataset contains 7,916 scenes and 11.27 billion 3D Gaussian splats. Constructing this dataset required computational resources equivalent to 150 GPU-days on one NVIDIA L4 GPU. SceneSplat-7K achieves high-fidelity reconstruction quality, with an average PSNR of 29.64 dB and a depth_l1 loss of 0.035 m.

The dataset was collected to train our feed-forward open-vocabulary 3DGS scene encoder, SceneSplat.

Please find here the links to each component of the dataset:

Data Statistics

We provide statistics for all the 3DGS scenes in the statistics folder. The statistics are saved in CSV files and include the psnr, ssim, lpips, depth_l1, and num_GS metrics. Using these statistics, users can filter the scenes by appearance and geometry quality.

The following table summarizes the lastest statistics after data updates.

Dataset	Scenes	Mean PSNR ↑	Mean SSIM ↑	Mean LPIPS ↓	Mean Depth L1 ↓	Mean #3DGS	Total #3DGS
ScanNet	1,613	30.17 dB	0.875	0.221	0.0151 m	1.000M	1.613B
ScanNet++	330	28.89 dB	0.917	0.144	0.0201 m	1.552M	512.282M
ScanNet++ v2	956	35.08 dB	0.959	0.062	0.0031 m	1.500M	1.434B
Replica	18	39.43 dB	0.978	0.056	0.0027 m	1.500M	27.000M
HyperSim	455	32.51 dB	0.942	0.078	0.0025 m	2.494M	1.135B
3RScan	632	27.06 dB	0.874	0.344	0.0177 m	1.500M	948.000M
ARKitScenes	1,290	31.63 dB	0.907	0.217	0.0051 m	1.149M	1.483B
Matterport3D	90	29.04 dB	0.882	0.208	0.0161 m	12.000M	1.080B
Total / mean	5,384	31.16 dB	0.906	0.189	0.0101 m	1.529M	8.232B

Frames Metadata

We first process the scenes and obtain per-scene metadata for all the training frames, which is saved in transforms_train.json. All the files are provided in the 3dgs_training_views folder. Each JSON file follows the format shown in the following example. Note that the poses use the OpenGL camera convention only for ScanNet++ scenes, as in the original dataset.

Contents of transforms_train.json

{
    "share_intrinsics": false,     // usually true; if false, use per-frame intrinsics in frames
    "fx": 0,                       // placeholder if share_intrinsics is false
    "fy": 0, 
    "cx": 0,  
    "cy": 0,  
    "width": width,                     // original image size
    "height": height,  
    "zipped": false,                    // if zipped, load later from ZIP files
    "crop_edge": 0,                     // crop the image edges if needed
    "resize": [960, 720],               // (width, height), resized for optimization
    "frames_num": "len(frames)",        // total number of frames for 3DGS optimization
    "init_point_num": "init_point_num", // point cloud size used for 3DGS initialization
    "bbox_min": "bbox_min",             // bounding box of the point clouds
    "bbox_max": "bbox_max",             
    "frames": frames,                   // see below
    "test_frames": test_frames          // if not provided by the dataset, randomly select 50

    // frames format: a list of dictionaries, each dictionary has the following format:
    [
        {
            "file_path": relative_path,           // image path relative to the scene folder
            "transform_matrix": transform_matrix, // 4 x 4, array.tolist(), camera-to-world pose
            "fx": intrinsics["fx"],               // needed if share_intrinsics is false
            "fy": intrinsics["fy"],
            "cx": intrinsics["cx"],
            "cy": intrinsics["cy"],
        },
        ...... // more frames
    ]
}

Preprocessed Language Pretraining Data

For convenience, we provide the preprocessed 3DGS vision-language pretraining data used for joint training of SceneSplat.

For each scene, the parameters from the 3DGS *.ply files are stored in separate *.npy files, and the 3DGS language labels are stored in lang_feat.npy and valid_feat_mask.npy.

.
├── color.npy
├── coord.npy
├── opacity.npy
├── quat.npy
├── scale.npy
├── lang_feat.npy
└── valid_feat_mask.npy

The following subfolders are required for vision-language pretraining:

scannet_mcmc_3dgs_lang_base: "train_grid1.0cm_chunk6x6_stride3x3", "test_grid1.0cm_chunk6x6_stride3x3", "val"
scannetpp_v2_mcmc_3dgs_lang_base: "train_grid1.0cm_chunk6x6_stride3x3", "test_grid1.0cm_chunk6x6_stride3x3", "val"
matterport3d_scene_mcmc_3dgs_lang_base: "train_grid1.0cm_chunk6x6x4_stride4x4x4", "val_grid1.0cm_chunk6x6x4_stride4x4x4", "valid_segment_splits", "test_eval"

2D Language Features

We provide the SigLIP2 2D language features extracted from the selected training frames, which are listed in each lang_feat_selected_imgs.json file in 3dgs_training_views. This process is detailed in the 3DGS Language Label Collection section of the main paper. The <frame_id>_f.npy file stores the per-frame SigLIP2 embeddings with shape (num_segs, 768), and the <frame_id>_s.npy file stores the per-frame SAM2 segmentation maps with shape (1, H, W).

Released 2D language features:

Additionally, we provide the 2D language features extracted using SigLIP2-so400m. The extraction process is the same, but uses the large SigLIP2-so400m model with a feature dimension of 1,152.

ScanNet, ScanNet++ V2, Matterport3D

Data Splits

For the benchmark results reported on the full evaluation scenes in the main paper, we use the official splits provided by each dataset: the ScanNet val split (312 scenes), the ScanNet++ nvs_sem_val split (50 scenes), and the Matterport3D test split (370 scenes). The split files are provided in the data_splits folder.

Dataset License

SceneSplat-7K is built upon multiple existing 3D datasets, each with its own licensing requirements. We've carefully structured our distribution approach to respect all original licenses while making our dataset accessible to the research community.

Distribution Approaches

Based on each dataset's licensing terms, we employ different distribution strategies:

Direct Distribution with Attribution: For datasets that permit redistribution for non-commercial research purposes (ARKitScenes, Hypersim, 3RScan, Replica), we include the original data with proper attribution and in accordance with the applicable license requirements.

Hosting Through Original Platforms: For datasets with custom terms that restrict redistribution, we've reached agreements with the original authors to host our processed 3D Gaussian Splatting scenes on their official platforms.

License Summary

Dataset	Original License	Allowed Purposes	Our Distribution Method
ARKitScenes	Apple Software License	Non-commercial use; modification and redistribution permitted	Direct distribution with attribution
Hypersim	CC BY-SA 3.0	Free to share and adapt under Attribution-ShareAlike terms	Direct distribution with attribution
3RScan	3RScan Terms of Use	Non-commercial research and educational purposes only	Direct distribution with attribution
Replica	Replica Dataset Research Terms	Non-commercial research or educational purposes	Direct distribution with attribution
ScanNet	ScanNet Terms of Use	Non-commercial research and educational purposes only	Hosted on original platform
ScanNet++	ScanNet++ Terms of Use	Non-commercial research and educational purposes only	Hosted on original platform
Matterport3D	Matterport License Agreement for Academic Use	Non-commercial academic use only	Direct distribution with attribution

Data Usage

When using the SceneSplat-7K dataset, please ensure that you agree to the following:

License Compatibility: All component datasets restrict use to non-commercial research and educational purposes.
Attribution: Proper attribution must be given to both SceneSplat-7K and all original dataset authors.
Data Access: For datasets hosted on original platforms (ScanNet, ScanNet++), you need to request access directly from those platforms.

The 3D Gaussian Splatting scenes we provide are governed by the original dataset licenses as detailed above. Our additional code, processing scripts, and metadata are made available under CC BY-SA 4.0.

If you have any questions about licensing, please reach out to us.

Citation

If you find our work helpful, please consider citing:

@inproceedings{li2025scenesplat,
  title={SceneSplat: Gaussian Splatting-based Scene Understanding With Vision-Language Pretraining},
  author={Li, Yue and Ma, Qi and Yang, Runyi and Li, Huapeng and Ma, Mengjiao and Ren, Bin and Popovic, Nikola and Sebe, Nicu and Konukoglu, Ender and Gevers, Theo and others},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2025}
}

Acknowledgements

We sincerely thank all the author teams of the original datasets for their contributions and for making their data publicly available. Our 3DGS scenes are optimized using the gsplat repository.