Prepare nuScenes-Occupancy

1. Download nuScenes V1.0 full dataset data HERE. Folder structure:

OpenOccupancy
├── data/
│   ├── nuscenes/
│   │   ├── maps/
│   │   ├── samples/
│   │   ├── sweeps/
│   │   ├── lidarseg/
│   │   ├── v1.0-test/
│   │   ├── v1.0-trainval/

2. Download the generated train/val pickle files and put them in data. Folder structure:

OpenOccupancy
├── data/
│   ├── nuscenes/
│   │   ├── maps/
│   │   ├── samples/
│   │   ├── sweeps/
│   │   ├── lidarseg/
│   │   ├── v1.0-test/
│   │   ├── v1.0-trainval/
│   │   ├── nuscenes_occ_infos_train.pkl/
│   │   ├── nuscenes_occ_infos_val.pkl/

2. Pre-compute depth map for fast training (depth-aware view transform module, same logic as BEVDepth):

python ./tools/gen_data/gen_depth_gt.py

Folder structure:

OpenOccupancy
├── data/
│   ├── nuscenes/
│   ├── depth_gt/

3. Download and unzip our annotation for nuScenes-Occupancy:

Subset	Google Drive	Baidu Cloud	Size
trainval-v0.0(deprecated)	link	link (code:BS95)	approx. 24G

Note that the v0.0 version is deprecated, and we release the v0.1 version, new features:

• Less occupancy noises, especially the artifacts caused by moving objects.
• More lightweight (V0.0: 24GB-before unzip, 270GB-after unzip. V0.1: 5GB-before unzip, 130GB-after unzip).
• Impreoved performance: v0.1 pretrained models enhance the mIoU by ~0.3 (compared to v0.0 pretrained models).

Subset	Google Drive	Baidu Cloud	Size
trainval-v0.1	link	link (code:25ue)	approx. 5G

We will release annotation (with more iterations of augmenting and purifying) in the future.

mv nuScenes-Occupancy-v0.1.7z ./data
cd ./data
7za x nuScenes-Occupancy-v0.1.7z
mv nuScenes-Occupancy-v0.1 nuScenes-Occupancy

Folder structure:

OpenOccupancy
├── data/
│   ├── nuscenes/
│   ├── depth_gt/
│   ├── nuScenes-Occupancy/

Basic information of nuScenes-Occupancy

*Note that we ignore noise, and set empty as label 0 in the training phase.