uild42/gate-gsplat-assets

gatescenesexport

Self-contained bundle of two sagesplat scenes (left/right) plus the object extractions in their shared (joint) mocap frame.

Layout

gate_scenes_export/
├── README.md
├── MANIFEST.txt
├── left_scene/
│   ├── mocap_processed/                         # data dir (referenced by config.yml as "./mocap_processed")
│   │   ├── images/                              # 242 corrected RGB frames
│   │   ├── transforms.json                      # mocap-frame camera poses
│   │   └── sparse_pc.ply                        # initial point cloud (mocap frame)
│   └── mocap_outputs/sagesplat_mocap/sagesplat/2026-05-11_153901/
│       ├── config.yml                           # trainer config (data path is relative)
│       ├── dataparser_transforms.json           # mocap → nerfstudio internal frame
│       └── nerfstudio_models/step-000029999.ckpt
├── right_scene/                                  # mirror layout
│   └── ...                                      #   training run timestamp 2026-05-11_144353
└── objects_final/
    ├── README.md                                # detailed transform usage
    ├── left_gate.ply,  right_gate.ply           # all PLYs in JOINT mocap frame
    ├── left_table.ply, right_table.ply
    ├── joint_mocap_to_nerf.json                 # the transforms
    ├── objects_summary.json                     # per-object AABBs, plane, polygon
    └── right_to_left_icp.json                   # ICP that built joint mocap

config.yml stores the data path as the relative string mocap_processed, so ns-viewer/ns-eval/ns-render must be launched with the scene directory as cwd. Everything else (checkpoint, dataparser transform) is referenced as a path relative to the run timestamp directory.

Environment

This bundle was produced with the conda env sagesplat at /home/javier/miniconda3/envs/sagesplat. To reproduce on another host:

• —python 3.10
• —pytorch 2.1.2 + cu118
• —nerfstudio 1.1.5
• —gsplat 1.4.0
• —tinycudann (built against the same CUDA toolkit)
• —sagesplat (editable install from Splat-MOVER)
• —runtime env knobs we use on this machine:
• —PYTHONNOUSERSITE=1 (so ~/.local/lib/python3.10/site-packages doesn't shadow the conda env)
• —QT_QPA_PLATFORM=offscreen (only relevant when running headless)
• —CUDA_HOME set to a conda env that ships a full CUDA toolkit, with LIBRARY_PATH extended with that env's stubs+system libcuda

Launching ns-viewer for a scene

# left
cd gate_scenes_export/left_scene
ns-viewer --load-config mocap_outputs/sagesplat_mocap/sagesplat/2026-05-11_153901/config.yml

# right
cd gate_scenes_export/right_scene
ns-viewer --load-config mocap_outputs/sagesplat_mocap/sagesplat/2026-05-11_144353/config.yml

Viser will bind to http://0.0.0.0:7007. ns-eval / ns-render take the same --load-config argument.

Using the object point clouds

See objects_final/README.md for the full transform chain and copy-pastable Python snippets. Short version:

• —All PLYs in objects_final/ are in the joint mocap frame: z-up, origin at the ArUco tag, +x along the tag's printed +y direction.
• —The right scene's PLYs have an ICP correction baked in so they align with the left scene's mocap (the joint reference).
• —objects_final/joint_mocap_to_nerf.json carries the per-scene 4×4 matrices needed to push points from joint mocap into either splat's nerfstudio-internal frame (for use with cv.aruco-style PnP, or to query the splat for renders at a specific mocap pose).

Frame conventions, quick reference

Total size: ~12 GB

Per-component:

• —left_scene/ : 76 MB data + 4.9 GB checkpoint = ~5.0 GB
• —right_scene/ : 71 MB data + 6.3 GB checkpoint = ~6.4 GB
• —objects_final/ : 11 MB

MANIFEST.txt carries a build-time snapshot of the same.