Datasets:
Tasks:
Feature Extraction
Modalities:
Image
Formats:
imagefolder
Languages:
English
Size:
< 1K
ArXiv:
License:
Demo data layout (data_demo/)
All demo cases read from this folder. Download from Hugging Face — PiLoT-data:
# from PiLoT repo root
./scripts/download_data_demo.sh
Partial downloads: ./scripts/download_data_demo.sh feicuiwan or smbu.
data_demo/
├── 3dgs_model/ # Feicuiwan / Jadebay PLY
│ └── point_cloud.ply
├── smbu_model/ # SMBU CityGaussian checkpoint + COLMAP sparse
│ ├── checkpoints/
│ │ └── epoch=60-step=30000.ckpt
│ └── sparse/sparse/0/
├── pretrained_model/ # PiLoT refiner weights (shared by all cases)
│ └── model@mapscape@512@Fourier.ckpt
└── query/
├── images/
│ ├── 3dgs_test/ # Feicuiwan query frames
│ └── smbu_seq2/ # SMBU query frames
└── poses/
├── 3dgs_test.txt # lon lat alt roll pitch yaw
└── smbu_seq2.txt # x y z pitch roll yaw (normalized)
Pose formats
| File | Columns |
|---|---|
3dgs_test.txt |
name lon lat alt roll pitch yaw |
smbu_seq2.txt |
name x y z pitch roll yaw |
Local development note
If disk space is tight, smbu_model/ may be a symlink to your CityGaussian
training output. The Hugging Face release ships real checkpoint files under
data_demo/smbu_model/.
| Variable | Default |
|---|---|
SMBU_MODEL_DIR |
data_demo/smbu_model |
CITYGAUSSIAN_ROOT |
third_party/CityGaussian |