Skip to content

Latest commit

 

History

History
66 lines (60 loc) · 2.42 KB

DATASET.MD

File metadata and controls

66 lines (60 loc) · 2.42 KB

Prepare Dataset

Folder structure

Drive-OccWorld
├── projects/
├── tools/
├── pretrained/
│   ├── r101_dcn_fcos3d_pretrain.pth
├── data/
│   ├── can_bus/
│   ├── nuscenes/
│   │   ├── maps/
│   │   ├── samples/
│   │   ├── sweeps/
│   │   ├── lidarseg/
│   │   ├── v1.0-test/
|   |   ├── v1.0-trainval/
|   |   ├── nuscenes_infos_temporal_train_new.pkl
|   |   ├── nuscenes_infos_temporal_val_new.pkl
│   ├── cam4docc
│   │   ├── GMO/
│   │   │   ├── segmentation/
│   │   │   ├── instance/
│   │   │   ├── flow/
│   │   ├── MMO/
│   │   │   ├── segmentation/
│   │   │   ├── instance/
│   │   │   ├── flow/
│   ├── nuScenes-Occupancy/

1. Download nuScenes V1.0 full dataset and CAN bus expansion data HERE. Prepare nuscenes data by running:

cd Drive-OccWorld
mkdir data
ln -s /path/to/nuscenes data/nuscenes
ln -s /path/to/nuscnens/can_bus data/can_bus

2. Prepare nuscene.pkl file

  1. Download the pre-processed nuscenes_infos_temporal_train_new.pkl and nuscenes_infos_temporal_val_new.pkl
  2. Move the .pkl files into the nuscenes folder:
mv nuscenes_infos_temporal_train_new.pkl /path/to/nuscenes
mv nuscenes_infos_temporal_val_new.pkl /path/to/nuscenes

3. Generate the inflated-occupancy annotations as defined in Cam4DOcc:

Create the dataset folders, which will automatically generate data during the first epoch of training and testing:

mkdir /path/to/save_data/cam4docc   # ~650G
cd /path/to/save_data/cam4docc
mkdir GMO GMO_lyft MMO MMO_lyft
ln -s /path/to/save_data/cam4docc Drive-OccWorld/data

4. Download the fine-grained occupancy annotations as defined in nuScenes-Occupancy. Then organize the dataset as follows:

mv nuScenes-Occupancy-v0.1.7z /path/to/save_data
cd /path/to/save_data
7za x nuScenes-Occupancy-v0.1.7z
mv nuScenes-Occupancy-v0.1 nuScenes-Occupancy
ln -s /path/to/save_data/nuScenes-Occupancy Drive-OccWorld/data/