--- license: cc-by-nc-4.0 pretty_name: Spatial 4D — Processed Gaussian Splatting Results tags: - gaussian-splatting - 4d-reconstruction - spatial4d - 3d --- # Spatial 4D — Processed Gaussian Splatting Results Experimental reconstructed 3D Gaussian sequences produced by Ryan Abraham's Spatial 4D project from the Neural 3D Video flame_salmon_1 scene. This release contains derived reconstruction outputs, not the original camera videos or a newly captured dataset. ## Contents `salmon-longseq-sequence-v2.zip` contains 31 binary PLY Gaussian states and a `spatial4d-sequence-v1` manifest: 155,304 persistent-index splats per state; source times 3.1333–8.1333 seconds; approximately 6 sampled states per second. The standalone `manifest.json` duplicates the archive manifest for inspection. Camera intrinsics and camera-to-world poses are included. Extract the ZIP and open its manifest in a compatible Spatial 4D sequence viewer. A generic PLY viewer can inspect individual states; it may not interpret Gaussian properties or sequence timing. ## Source and credits Original capture: **Neural 3D Video Synthesis from Multi-view Video**, Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv (CVPR 2022). - Original dataset: https://github.com/facebookresearch/Neural_3D_Video - Paper: https://arxiv.org/abs/2103.02597 - Original project: https://neural-3d-video.github.io/ Changes: multiview video preprocessing, Gaussian reconstruction and temporal fitting, and export to persistent-index PLY states with viewer metadata. Reconstruction and release packaging: Ryan Abraham. Original capture credit remains with the source authors. No endorsement by those authors is implied. ## License CC BY-NC 4.0. Preserve attribution and noncommercial restrictions. The upstream license is included in `LICENSE`; https://creativecommons.org/licenses/by-nc/4.0/ . ## Limitations Experimental sequential reconstruction, not a validated continuous deformation model. Wide viewpoints remain unvalidated; reconstruction artifacts and temporal inconsistencies may occur. Calibration was reused from a separate single-timestamp 960×720 calibration preparation and was not re-derived for this sequence. cam00 was used for validation/model selection and is not an untouched test camera. No benchmark scores are asserted for this release.