Motoki098/tennis-lab-multiview-sample

日本語

Tennis-Lab Multiview Sample

These are short tennis videos from Meiji, recorded with three synchronized fixed cameras, for trying Tennis-Lab's multiview 3D reconstruction. The sample covers the entire exported video_000/clip_000. Consent has been obtained from the people filmed for the footage being published (project agreement record). Raw videos, training annotations, annotations/, outsource/, inference results, and model weights are not included.

Code commit: cb8f1be66c34256046d1388962ee0f1dde3a9811

Files and input requirements

cam0.mp4      # 11850811 bytes
cam1.mp4      # 19479451 bytes
cam2.mp4      # 19787313 bytes
clip.json     # 1574 bytes
README.md
README.ja.md
LICENSE       # 19127 bytes
SHA256SUMS

All three videos are 1920×1080 at 59.94006 fps, with 1,010 frames (about 16.8502 seconds). The exported files are H.264 / yuv420p MP4s; no additional trimming or re-encoding was performed. The capture date is not recorded in the manifest. Details of the venue and equipment, and any remaining synchronization error, have not been verified.

clip.json is the public manifest and references the videos by paths relative to this directory. The three local absolute source_path values in the original export manifest were replaced with relative paths identifying the raw videos, and video_paths and each camera's video were changed from media/cam*.mp4 to cam*.mp4. All other values, including the timeline, synchronization values, source frame ranges, and letterbox information, were retained. source_path identifies provenance; it is not a raw-video download location.

CPU checks with build_clip_source() and the standard pipeline's sampled_frame_indices() confirmed three views with matching FPS, resolution, and frame counts, and a grid of 506 frames at approximately 30 fps. This is within frame_sampling.max_frames=1024; the first and last grid entries are source video frames 0 and 1009. Every frame of all three videos was also decoded successfully on CPU. The standard pipeline was additionally confirmed to complete through saving a SceneResult v2 on a GPU environment with this sample (1,010 frames, status=ok); see the completion record.

cam0 letterbox

The cam0 raw video is 1280×960; the cam1 and cam2 raw videos are 1920×1080. cam0 was scaled by 1.125 to 1440×1080 while preserving its aspect ratio, with 240 pixels of padding added to each side. All three final videos are 1920×1080. cam1 and cam2 have no letterboxing. Upscaling does not add detail to the original cam0 footage. The pipeline detects and calibrates in the exported pixel coordinates, so this layout satisfies its common-image-size input requirement. No preprocessing to remove the padding or restore the original resolution is needed.

Provenance and synchronization

The raw-data identifiers are tennis_multivew/raw/meiji_3cam/video_000/cam{0,1,2}.mp4. No new Clip Studio editing, synchronization calculation, or export was performed for this publication. The existing clip.json records sync_source=clip_studio and exported_at=2026-07-13T23:43:53.467690+00:00. This is the export time, not the capture time.

The saved project is projects.video_000 in tennis_multivew/processed/meiji_3cam/projects.json; the original export manifest is tennis_multivew/processed/meiji_3cam/dataset/videos/video_000/clips/clip_000/clip.json. During publication preparation, the saved project's clip interval, camera IDs, raw-video identifiers, and offsets were checked against the manifest and matched. The saved project does not establish how the offsets were estimated.

Camera	Raw resolution	Source offset (seconds)	Source frame range (inclusive)
cam0	1280×960	0.0	2238–3247
cam1	1920×1080	48.465083333332295	5143–6152
cam2	1920×1080	-31.998633333333203	320–1329

All raw videos have FPS 60000/1001 (about 59.94005994). The global time interval is 37.338406601390844–54.18857326805711 seconds; source time is global time + offset_sec. Source frame numbers are zero-based. The published MP4s use frame 0 after extraction as their common start, so do not apply the offsets again when using them.

The original export manifest and the entire saved project at verification time were identified by SHA-256. Those original files are not included in the public bundle.

Download and placement

After setting up Tennis-Lab, run these commands from the repository root. The CLI installs the three videos, public manifest, and LICENSE under data/samples/ with SHA-256 verification. Matching existing files are reused; mismatches or network failures are errors. Existing files are not overwritten.

.venv/bin/python -m src.utils.sample_dataset --data-root "$PWD/data"
.venv/bin/python -m src.utils.sample_dataset --data-root "$PWD/data" --verify-only

To pin the dataset version, add --revision <this dataset's Hub commit>. The code commit and the dataset's Hub commit are different. You can use the default video_paths=[samples/cam0.mp4,samples/cam1.mp4,samples/cam2.mp4] and camera_ids=[cam0,cam1,cam2] settings. See the Tennis-Lab installation guide for required model assets and running the sample. This dataset does not provide model weights.

If downloading the whole repository with the HF CLI, use a separate empty location outside Tennis-Lab's input directory and verify it after downloading. hf download is not Tennis-Lab's placement CLI and does not itself guarantee protection of existing files.

hf download Motoki098/tennis-lab-multiview-sample --repo-type dataset --local-dir /absolute/path/to/new/sample-bundle
cd /absolute/path/to/new/sample-bundle
sha256sum -c SHA256SUMS

SHA-256

Checksums for the three videos, clip.json, and LICENSE are listed in SHA256SUMS, and the download CLI independently verifies these files against the values pinned in its code.

Terms of use

The videos and public manifest are published under CC BY-NC-ND 4.0. You may share the unmodified material noncommercially, subject to attribution and the other license terms. You may create and reproduce adaptations for noncommercial purposes, but sharing them is not permitted. If crops, annotation overlays, GIFs, or similar changes constitute adaptations, sharing them requires separate permission. Necessary technical format changes and uses that do not require permission under applicable law are governed by the license text.

Attribution example: Motoki0705, “Tennis-Lab Multiview Sample”, 2026, distribution page, CC BY-NC-ND 4.0.

The intended uses are research and noncommercial demos; this description adds no use restrictions. CC does not uniformly grant rights such as likeness or privacy rights. Separately used assets such as GVHMR, SMPL, and SMPL-X, and pipeline outputs, remain subject to their respective terms. This dataset's terms do not apply collectively to Tennis-Lab's original code or third-party model assets. See the project's sample dataset terms and pipeline output terms. For the full text, see the included LICENSE. The LICENSE preserves the official Creative Commons text unchanged.