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