README.md

Data directory

The pipeline reads its inputs from and writes its outputs under this data/ directory (set by paths.data_root in each config; the shipped configs use ../data). The large inputs are not stored in git — this folder is .gitignored except for this README.

Expected layout

data/
├── stl_templates/                       # copy or symlink of ../assets/stl_templates/
│   ├── chair_3.stl  table_2.stl  person_sitting_1.stl  ...
└── projects/<name>/
    └── input/
        ├── point_cloud_cropped.ply       # dense NeRF-exported point cloud (GBs)
        ├── transforms.json               # camera intrinsics + poses
        ├── dataparser_transforms.json    # Nerfstudio dataparser transform
        └── frames/
            ├── chair/*.jpg               # frames for the "chair" category
            └── big_table/*.jpg           # frames for the "table" category

Outputs are written to data/projects/<name>/runs/<run_id>/ (see docs/pipeline.md).

Project names for the three paper environments

Config	project.name
configs/classroom_a.yaml	classroom_v2_v2
configs/classroom_b.yaml	classroom_v1
configs/auditorium.yaml	audi_v1

Getting the input data

Set up the stl_templates/ folder (copy or symlink ../assets/stl_templates/), then place each environment's input/ under data/projects/<name>/.

The reconstruction inputs (NeRF point clouds, camera transforms, and frames; ~7 GB total for the three environments) are available from the authors on reasonable request.

If you generate your own data, produce point_cloud_cropped.ply by training a Nerfstudio nerfacto model and running ns-export pointcloud, and place the matching transforms.json / dataparser_transforms.json and per-category frames as shown above.