louenpottier/lagsplat-mpov

A four-view smartphone video dataset of a robot-held flexible strip, with camera poses and a 3D LaGSplat baseline

Canonical record and DOI: 10.5281/zenodo.23601189 (Zenodo). This Hugging Face repository mirrors the Zenodo record file for file, with the archives unpacked so that each file can be browsed and downloaded on its own.

• Dataset note (setup, synchronization, calibration, file formats, results): mpov_dataset_note.pdf
• Training code: https://github.com/LouenPottier/lagsplat-mpov
• Real-time 3D demo in the browser: https://louenpottier.github.io/xrdays.html
• Method paper: LaGSplat, arXiv:2608.16324

Summary

Four synchronized smartphone videos of a robot arm holding a flexible strip (about 23 cm long), with the measured pose of every camera. The arm moves by jerks along a nearly one-dimensional horizontal path; after each jerk, the strip oscillates freely and its motion decays. Three iPhones (iphone0, iphone1, iphone2) are fixed and mounted rigidly to one another; a fourth (iphone3) is handheld and moves around the scene for the first seven minutes, then is held nearly still.

All videos are 1920 × 1080 at 60000/1001 fps, resampled on a common clock (that of iphone1) so that frame k of every video shows the same instant, with a residual of at most half a frame (8 ms). Each video has 36 734 frames (612.8 s). Two further clips, filmed with every object in the room still, give a static reconstruction of the room.

The world frame is the camera frame of iphone0 (OpenCV convention); the unit of length is the distance between iphone0 and iphone1 (about 0.38 m). No calibration target was filmed: the fixed cameras are solved by a joint bundle adjustment on clicked correspondences (reprojection error of 3 to 6 px), the handheld camera by PnP against a SIFT map (25 381 of the 36 734 frames have a pose), the static clips by COLMAP aligned to the world frame.

The record also ships the checkpoints of a 3D baseline: LaGSplat fitted on this footage, where the strip is rendered by Gaussian Splatting, driven by a learned Lagrangian with the acceleration of the gripper as forcing term, and responds in real time to forces applied by a user. On the held-out last 60 s, rendered images reach a PSNR of 31.4 to 32.8 dB for the fixed cameras and 26.5 dB for the handheld one.

Files

Path	Content	Size
scene_mobile/iphone{0,1,2,3}.mp4	synchronized session, H.264, 36 734 frames each, no audio	2.0 G, 382 M, 344 M, 590 M
scene_mobile/frames.csv	per frame k: common time, source frame of each camera, residual offset	3.6 M
scene_mobile/sync.json	cut, clock rates, dropped frames, checks
scene_statique/iphone3_fixed_scene{,_2}.mp4	static room, 1920 × 1080, 30 fps, 6 303 and 1 722 frames	229 M, 42 M
scene_statique/sync.json	metadata of the static clips
geometry/cameras_fixed.json	intrinsics, distortion and pose of iphone0, 1, 2
geometry/iphone3_poses.csv, geometry/iphone3_intrinsics.json	pose of iphone3 per frame k, and its intrinsics
geometry/static_poses.json	poses of the static clips, in the same world frame
geometry/landmarks.json, geometry/zones.json	3D static landmarks; oriented boxes of the zones s (moving part of the arm) and q (strip)
geometry/state.csv	per frame k: measured input s(t) of the arm and baseline latent state q(t) ∈ ℝ²
runs/	checkpoints of the 3D baseline (contents of checkpoints.zip on Zenodo): static scene, autoencoder, decoder, Lagrangian neural network, linear dynamics	74 M
mpov_dataset_note.pdf	the dataset note

Correspondence with the Zenodo archives: scene_mobile.zip → scene_mobile/, scene_statique.zip → scene_statique/, geometry.zip → geometry/, checkpoints.zip → runs/.

Limitations

The calibration error is a few pixels, not sub-pixel, and some poses of iphone3 are wrong during the handheld part. s(t) is read in one view, in pixels. The colours of iphone0 differ from the other cameras (it is a different model).

License

CC BY-NC 4.0.

Citation

Please cite the dataset note when using the data:

@misc{pottier2026mpov,
  title     = {A four-view smartphone video dataset of a robot-held flexible strip,
               with camera poses and a 3D LaGSplat baseline},
  author    = {Pottier, Louen},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23601189},
  url       = {https://doi.org/10.5281/zenodo.23601189}
}

and the method paper:

@misc{pottier2026lagsplat,
  title         = {LaGSplat: Inferring Physics-Governed Interactive Simulation from
                   Monocular Video Using Latent Lagrangian Gaussian Splatting},
  author        = {Pottier, Louen},
  year          = {2026},
  eprint        = {2608.16324},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2608.16324}
}