Skip to content

Latest commit

 

History

History
62 lines (59 loc) · 2.78 KB

prepare_data.md

File metadata and controls

62 lines (59 loc) · 2.78 KB
Raw

Prepare nuScenes-H

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

ASAP
├── data/
│   ├── nuscenes/
│   │   ├── maps/
│   │   ├── samples/
│   │   ├── sweeps/
│   │   ├── v1.0-test/
|   |   ├── v1.0-trainval/

Generate 12Hz annotations: For convinience, the 12Hz annotations can be simiply calculated by the object interpolation:

bash scripts/ann_generator.sh 12 --ann_strategy 'interp'

Or you can use the advanced method (object interpolation + temporal database) to generate 12Hz annotation. Firstly you should train CenterPoint on the trainval set of nuScnese (follow the official instructions in MMDetection3D, and we provide config files in ./asset/centerpoint_0075voxel_second_secfpn_dcn_circlenms_4x8_cyclic_20e_nus_trainval.py). Then we use the pretrained CenterPoint to generate detection results for 20Hz LiDAR input:

  1. Generate 20Hz LiDAR input pkl file for CenterPoint
bash scripts/nusc_20Hz_lidar_input_pkl.sh
  1. Generate 20Hz detection results. We provide a template inference script for CenterPoint inference (use MMDetection3D):
python tools/test.py \
    $PATH_TO_ASAP/assets/centerpoint_20Hz_lidar_input.py \
    $pretrained_model_path (use the above)\
    --eval-options 'jsonfile_prefix=$PATH_TO_MMDetection3D/work_dirs/centerpoint_0075voxel_second_secfpn_dcn_circlenms_4x8_cyclic_20e_nus_trainval/' \
    --out $PATH_TO_MMDetection3D/work_dirs/centerpoint_0075voxel_second_secfpn_dcn_circlenms_4x8_cyclic_20e_nus_trainval/rst.pkl \
    --format-only
  1. Consequently, we obtain 20Hz inference results at $PATH_TO_MMDetection3D/work_dirs/centerpoint_0075voxel_second_secfpn_dcn_circlenms_4x8_cyclic_20e_nus_trainval/pts_bbox/results_nusc.json. Then we build the temporal database:
bash scripts/nusc_20Hzlidar_instance-token_generator.sh --lidar_inf_rst_path $PATH_TO_MMDetection3D/work_dirs/centerpoint_0075voxel_second_secfpn_dcn_circlenms_4x8_cyclic_20e_nus_trainval/pts_bbox/results_nusc.json
  1. Finaly, generate the 12Hz annotation:
bash scripts/ann_generator.sh 12 \
   --ann_strategy 'advanced' \
   --lidar_inf_rst_path ./out/lidar_20Hz/results_nusc_with_instance_token.json

After the data preparation , the folder structure is:

ASAP
├── data/
│   ├── nuscenes/
│   │   ├── maps/
│   │   ├── samples/
│   │   ├── sweeps/
│   │   ├── v1.0-test/
|   |   ├── v1.0-trainval/
|   |   ├── interp_12Hz_trainval/
|   |   ├── advanced_12Hz_trainval/ (optional)
  1. Visualize the 12Hz annotation.
bash scripts/render_ann.sh interp_12Hz_trainval --render_anns