Skip to content

Latest commit

 

History

History
62 lines (48 loc) · 2.25 KB

DATA_PREP.md

File metadata and controls

62 lines (48 loc) · 2.25 KB

Prepare Bench2Drive Dataset

Download Bench2Drive

Download Bench2Drive dataset from (LINK) and make sure the structure of data as follows:

Notice: some version of data may have slightly different folder structure. You may need to use soft link (ln -s) and change the path related code.

    Bench2DriveZoo
    ├── ...                   
    ├── data/
    |   ├── bench2drive/
    |   |   ├── v1/                                        # Bench2Drive base 
    |   |   |   ├── Accident_Town03_Route101_Weather23/
    |   |   |   ├── Accident_Town03_Route102_Weather20/
    |   |   |   └── ...
    |   |   └── maps/                                        # maps of Towns
    |   |       ├── Town01_HD_map.npz
    |   |       ├── Town02_HD_map.npz
    |   |       └── ...
    |   └── splits
    |           └── bench2drive_base_train_val_split.json    # trainval_split of Bench2Drive base

Preprocess Bench2Drive data info

Run the following command to preprocess data:

cd adzoo/ofae2e/mmdet3d_plugin/datasets
python preprocess_bench2drive_drivetransformer.py --workers 16

The structure of generated data is:

    Bench2DriveZoo
    ├── ...
    ├── data/
    |   ├── bench2drive/
    │   ├── infos/
    │   │   ├── b2d_infos_v1_train_drivetransformer/
    |   |   |   ├── Accident_Town03_Route101_Weather23.pkl
    |   |   |   ├── Accident_Town03_Route102_Weather20.pkl
    |   |   |   └── ...
    │   │   ├── b2d_infos_v1_val_drivetransformer/
    |   |   |   ├── Accident_Town05_Route218_Weather10.pkl
    |   |   |   ├── AccidentTwoWays_Town12_Route1115_Weather23.pkl
    |   |   |   └── ...
    |   |   |—— b2d_infos_v1_train_drivetransformer_meta.pkl
    │   │   ├── b2d_infos_v1_val_drivetransformer_meta.pkl
    |   |   └── b2d_map_infos.pkl
    |   └── splits/
    ├── ...

Note: This command will be by default use all routes except those in data/splits/bench2drive_base_train_val_split.json as the training set. It will take about 1-2 hours to generate all the data with 16 workers for Base set (1000 clips).