data_preparation.md

You can download nuScenes 3D detection data HERE and unzip all zip files. You can only download v1.0-trainval version conveniently. Like the general way to prepare dataset, it is recommended to symlink the dataset root to $SOGDet/data.

We typically need to organize the useful data information with a .pkl. To prepare these files for nuScenes, run the following command:

python tools/create_data.py  
python tools/data_converter/prepare_nuscenes_for_bevdet4d.py  

Then download nuscenes panoptic dataset HERE. Symlink the dataset root to $SOGDet/data/nuscenes and replace .json file following HERE. Run the following command to generate point cloud with panoptic labels.

python tools/data_converter/prepare_panoptic.py

Here, the total data structure is shown like this:

mmdetection3d
├── mmdet3d
├── tools
├── configs
├── data
│   ├── nuscenes
│   │   ├── maps
│   │   ├── panoptic
│   │   ├── pc_panoptic
│   │   ├── samples
│   │   ├── sweeps
│   │   ├── v1.0-test
|   |   ├── v1.0-trainval
│   │   ├── nuscenes_infos_train.pkl
│   │   ├── nuscenes_infos_val.pkl
│   │   ├── nuscenes_infos_train_4d_interval3_max60.pkl
│   │   ├── nuscenes_infos_val_4d_interval3_max60.pkl