prepare_data.md

Prepare the data

Dowload the nuScenes dataset

Download the nuScenes dataset and create symbolic links.

cd ${duocc_path}
mkdir data
ln -s path/to/nuscenes ./data

Dowload the Occ3D-nuScenes dataset

For Occupancy Prediction task, download Occ3D-nuScenes from CVPR2023-3D-Occupancy-Prediction and place it in data/nuscenes/gts

Prepare pkl files

Pack the meta-information and labels of the dataset, and generate the required .pkl files.

python3 tools/create_data_duocc.py

Generate anchors by K-means

python3 tools/anchor_generator.py --ann_file ./data/nuscenes_anno_pkls/nuscenes_occ_infos_aug_train.pkl

Download pre-trained weights

Download the required backbone BEVDet serires and place it in ckpts folder.

The Overall Structure

Please make sure the structure of DuOcc is as follows:

DuOcc
├── projects/
├── tools/
├── ckpts/
│   ├── bevdet-r50-4d-depth-cbgs.pth
├── data/
│   ├── nuscenes/
│   │   ├── maps/
│   │   ├── samples/
│   │   ├── samples/
│   │   ├── sweeps/
│   │   ├── v1.0-test/
│   │   ├── v1.0-trainval/
│   │   ├── nusceness_occ_infos_train.pkl
│   │   ├── nusceness_occ_infos_val.pkl
│   ├── nuscenes_anno_pkls/
│   │   ├── nuscenes_occ_infos_aug_train.pkl
│   │   ├── nuscenes_occ_infos_aug_val.pkl
├── nuscenes_kmeans900_42m.npy
└── others

→ Back to: Training & Inference