The scene.init checks for data format markers in order:
- sparse/ → COLMAP (static)
- transforms_train.json → Blender
- poses_bounds.npy → plenopticVideo (dynamic)
- ...etc.
Since the neur3d zip contains sparse/0/ (COLMAP reconstruction), and the COLMAP check comes first, the fast-d3dgs branch loads neur3d as static COLMAP data.
The train.sh confirms this — the training command is just:
python train.py -s ./data/neur3d/cook_spinach -m output/cook_spinach --eval --iterations 30000
No --is_blender, no --is_6dof, no temporal flags. Same format as regular FastGS static training.
So the fast-d3dgs branch treats neur3d scenes as static multi-view scenes (via COLMAP loader), even though the training pipeline still includes a deformation network. The deformation model is present but learns minimal/zero deformations since the data has no temporal axis.
The scene.init checks for data format markers in order:
Since the neur3d zip contains sparse/0/ (COLMAP reconstruction), and the COLMAP check comes first, the fast-d3dgs branch loads neur3d as static COLMAP data.
The train.sh confirms this — the training command is just:
python train.py -s ./data/neur3d/cook_spinach -m output/cook_spinach --eval --iterations 30000
No --is_blender, no --is_6dof, no temporal flags. Same format as regular FastGS static training.
So the fast-d3dgs branch treats neur3d scenes as static multi-view scenes (via COLMAP loader), even though the training pipeline still includes a deformation network. The deformation model is present but learns minimal/zero deformations since the data has no temporal axis.