Download nuScenes dataset here. Set a symbolic link to ./data/nuscenes.
Or if you are using docker installation, a shared volume is created here.
We follow MUTR3D to create nuScenes info files for tracking task, which is based on the mmdetection3d data pre-processing pipeline.
Simply run
python ./tools/data_converter/nusc_track.py
This script will create info files for all train/val/test splits.
The structure of the nuScenes dataset folder should look like.
nuscenes/
├──── v1.0-test
├──── v1.0-trainval
├──── maps/
├──── samples/
├──── sweeps/
├──── ada_track_infos_train.pkl
├──── ada_track_infos_val.pkl
└──── ada_track_infos_test.pkl
We build our track based on DETR3D and PETR. You can download following pretrained models from their repository:
- DETR3D ResNet101 from DETR3D
- PETR VoVNet (900 queries) from PETR.
Our PETR-based tracker uses 500 queries but the pretrained PETR model uses 900 queriey.
Fortunately, PF-Track provided a pretrained PETR model with 500 queries.
You can download their pre-trained model following their instruction.
Decompress the zip file and the file f1_q5_fullres_e24.pth is the checkpoint what we need.
Many thanks to aforemetioned open-source projects with their provided pre-trained models!