In order to train and evaluate the model, nuScenes, a large-scale autonomous driving dataset with 3D object annotations, is used. It consists of 1000 scenes captured from four locations in Boston and Singapore, each of 20 seconds in length, covering different conditions. The images are captured from 6 surround-view cameras which provides a 360° view with a slight overlap between the neighboring cameras. As the dataset already provides annotated 3D objects with a category, attributes and 3D bounding box it can be used for training and testing. See more detail HERE.
Download nuScenes V1.0 full dataset data and CAN bus expansion data HERE.
To save exporatation time, we have made the compressed blobs of nuScenes V1.0 full dataset available on SCITAS at /work/vita/datasets/bev/data/nuscenes_full_compressed
Download CAN bus expansion
# download 'can_bus.zip'
unzip can_bus.zip
# move can_bus to data dir
Prepare nuScenes data
BEVFormer repo generates custom annotation files which are different from mmdet3d's
python tools/prepare_data.py nuscenes --version v1.0 --root-path ./data/nuscenes --out-dir ./data/nuscenes --canbus ./data
Using the above code will generate nuscenes_infos_temporal_{train,val}.pkl.
To save time, we have made nuscenes_infos_temporal_{train,val}.pkl available HERE. To use them, place them according to the folder structure bellow and update the file with your data_root_path by running:
python update_nuscenes_infos.py data/nuscenes/nuscenes_infos_temporal_train.pkl data_root_path
python update_nuscenes_infos.py data/nuscenes/nuscenes_infos_temporal_val.pkl data_root_path
Repo architecture
BEVFormer_HoP_gr3
├── projects/
├── tools/
├── configs/
├── ckpts/
├── data/
│ ├── can_bus/
│ ├── nuscenes/
│ │ ├── maps/
│ │ ├── samples/
│ │ ├── sweeps/
│ │ ├── v1.0-mini/
│ │ ├── v1.0-test/
| | ├── v1.0-trainval/
| | ├── nuscenes_infos_temporal_train.pkl
| | ├── nuscenes_infos_temporal_val.pkl
Pretrained weights (ckpts/) can be downloaded HERE and added to the repo according to the repo architecture above.