Skip to content

Latest commit

 

History

History
39 lines (24 loc) · 2.19 KB

preprocessing.md

File metadata and controls

39 lines (24 loc) · 2.19 KB

Preprocessing nuScenes

1. Download nuScenes

Download the nuScenes dataset, soft-link it to ./data/nuscenes. This step is compulsory.

2. Creating infos file

We follow the practice in MMDetection3D and create information files for training/evaluation the tracking and motion prediction tasks for PF-Track. You can either use our provided files or optionally create the infos file on your own.

2.1 Using Our Provided Infos File.

If you haven't downloaded the provided files, checkout "pretrained models and data files." If you have finished downloading, copy the infos file into ./data/nuscenes/.

2.2 Reproducing the infos files. (Optional)

For training and validation splits, commands below generate tracking_forecasting_infos_train.pkl and tracking_forecasting_infos_val.pkl. (Around 3 hours)

python tools/create_data.py nuscenes-tracking --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag tracking_forecasting --version v1.0 --forecasting

For mini splits, commands below generate tracking_forecasting-mini_infos_train.pkl and tracking_forecasting-mini_infos_val.pkl. (Around 3 minutes)

python tools/create_data.py nuscenes-tracking --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag tracking_forecasting --version v1.0-mini --forecasting

For the test split, commands below generate tracking_forecasting_infos_test.pkl.

python tools/create_data.py nuscenes-tracking-test --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag tracking_forecasting --version v1.0 --forecasting

3. Notes for Developers

  • Core code? We change detection-style information file (mmdetection3d) into a tracking-style information file (PF-track) mainly from the following aspects. [Related Code 1][Related Code 2][Related Code 3].
  • How to support tracking? We load the ids of objects into the information files to form tracks across frames.
  • How to support prediction? We load the future trajectories (13 frames, or ~6.0 seconds) to support the training of motion prediction.