Datasets:

Modalities:
Video
Size:
< 1K
Libraries:
Datasets
License:

Accept the dataset evaluation terms

These recordings are for private, noncommercial evaluation and testing only. Read the LICENSE before requesting access. Commercial use and sharing footage, images, renders, reconstructions, or scene-specific models require prior written permission from Arizona State University Narrative and Emerging Media.

Log in or Sign Up to review the conditions and access this dataset content.

Open4DGS-Datasets

Multi-camera video captures for evaluating and testing the Open4DGS-Core preparation pipeline and the Open4DGS-Trigs trainer.

Evaluation and testing only. These datasets use a separate, restricted Dataset Evaluation License. They are not covered by the MIT or AGPL licenses of the software repositories. Noncommercial, private evaluation is permitted, including frame extraction and scene-specific reconstruction training. Sharing footage or derived visual/model assets requires prior written permission. See the license for the full terms.

Captures

Dataset Content Cameras Files per camera Total videos Approximate size
apt-dance-01 Dance capture 27 C0004.MP4, C0008.MP4 54 6.35 GB
spidermen-01 Two performers in Spider-Man costumes 45 C0038.MP4, C0046.MP4 90 11.00 GB

Sizes are decimal GB for the video files before archiving. All supplied videos are 1920 Γ— 1080 at 60000/1001 fps (approximately 59.94 fps) and contain embedded timecode. Each camera folder contains two separate takes; do not concatenate them or mix take numbers across cameras.

Dataset Performance take Alignment footage
apt-dance-01 C0004.MP4 (--take 4) C0008.MP4 (--take 8)
spidermen-01 C0038.MP4 (--take 38) C0046.MP4 (--take 46)

Use the alignment footage to prepare camera alignment, and the performance take for the dynamic reconstruction. Extract them into separate directories using a timecode interval appropriate to each take.

The supplied files are original camera video inputs. Extracted frames, camera calibration, refined timing estimates, masks, and trained reconstructions are not included in these folders.

Folder layout

Preserve the camera folder names, filenames, and embedded metadata when copying or unpacking the videos:

open4dgs-datasets/
β”œβ”€β”€ README.md
β”œβ”€β”€ LICENSE
β”œβ”€β”€ apt-dance-01/
β”‚   β”œβ”€β”€ Camera 01/
β”‚   β”‚   β”œβ”€β”€ C0004.MP4
β”‚   β”‚   └── C0008.MP4
β”‚   └── ...
└── spidermen-01/
    β”œβ”€β”€ Camera 01/
    β”‚   β”œβ”€β”€ C0038.MP4
    β”‚   └── C0046.MP4
    └── ...

Camera numbering is not necessarily contiguous. Keep the complete camera set for each take and avoid renaming or re-encoding the source recordings.

Download

The videos are hosted in the Hugging Face dataset repository. Sign in there and accept the evaluation terms to obtain access. Access is approved automatically after acceptance. The download preserves the dataset and camera folders; no archives or filename conversion are required.

Install the download tool and sign in with the same Hugging Face account:

python -m pip install --upgrade huggingface_hub
hf auth login

Download both captures (approximately 17.35 GB of video):

hf download asu-nem-labs/open4dgs-datasets \
  --repo-type dataset --local-dir ./open4dgs-datasets

Or download only one capture, together with its documentation and checksums:

hf download asu-nem-labs/open4dgs-datasets \
  --repo-type dataset --local-dir ./open4dgs-datasets \
  --include 'apt-dance-01/*' --include README.md --include LICENSE \
  --include 'checksums/apt-dance-01.sha256'

For the other capture, replace both occurrences of apt-dance-01 with spidermen-01. Re-run the same command to finish an interrupted download. Keep enough free disk space for the videos and any subsequently extracted frames. Extraction and training need additional storage.

To verify downloaded videos on Linux, run from the download directory:

sha256sum -c checksums/apt-dance-01.sha256
sha256sum -c checksums/spidermen-01.sha256

Check only the capture(s) you downloaded. On macOS, use shasum -a 256 -c in place of sha256sum -c.

Using the videos

  1. Install the tools using the Open4DGS setup guide.
  2. Select a dataset and a single take. Choose a common start timecode and a frame count that fit within every selected camera's recording. File start timecodes can differ; selecting the same file-relative frame number does not necessarily select the same moment across cameras.
  3. Extract frames using the embedded timecode, then follow the preparation and training pipeline.

The following is a command template, not a fixed benchmark interval. Replace the paths, start timecode, and frame count before running it. This example selects the apt-dance-01 performance take:

python /path/to/open4dgs-core/prep/extract/extract_synced_sequence.py \
  --input-dir /path/to/open4dgs-datasets/apt-dance-01 \
  --output-dir /path/to/work/apt-dance-01/images \
  --take 4 \
  --start-timecode <HH:MM:SS:FF> \
  --num-frames <FRAME_COUNT> \
  --skip-existing

For spidermen-01, change the input/output paths and use --take 38 for the performance. For alignment footage, use --take 8 or --take 46 respectively, with a separate output directory and its own extraction interval. Choose rotation settings appropriate to the footage; see the extraction tool's --help. Use separate output directories for different takes or extraction settings.

Embedded timecode supplies the extraction reference. Camera alignment is a later preparation step; visual synchronization can estimate residual sub-frame timing differences. Timecode alone is not a claim of hardware genlock. Preserve the actual 60000/1001 frame rate in downstream configuration rather than silently substituting 60 fps.

For reproducible comparisons, record the dataset, take, camera selection, start timecode, number of frames, rotations/resizing, core and trainer commit IDs, and preparation/training settings. No fixed evaluation split or benchmark scores are defined here yet.

Results and permissions

You may report aggregate numerical evaluation results with acknowledgment and a link to this repository. Permission is required before posting source footage, extracted images, screenshots, renders, or reconstructed models, including in demos, papers, and social posts. Linking others to the official dataset page is welcome.

The dataset rights holder is Arizona State University Narrative and Emerging Media. For questions or permission requests, contact atehran4@asu.edu. Do not attach dataset media to public issues without permission.

Third-party characters and trademarks remain the property of their respective owners. Their appearance does not imply affiliation or endorsement, and this dataset license does not grant rights to those characters or trademarks.

Downloads last month
77