Download the nuScenes dataset, soft-link it to ./data/nuscenes. This step is compulsory.
2. Download Preprocessed nuScenes Files [Here]
The generation of preprocessed nuScenes files can be found in PF-Track.
3. Download NuPrompt Dataset [Here]
Each json file comprises four parts:
{
'scene_token': <str> -- unique identifier of a scene
'prompt': <str> -- language prompt
'frame_token_object_token':{
'frame_token_1':[ <str> -- unique identifier of a frame
'instance_token_1', <str> -- unique identifier of an object
'instance_token_2',
...
]
'frame_token_2':[
'instance_token_1',
'instance_token_3',
...
]
...
}
'original_prompt': <str> -- original language prompt
},
Different from the original nuScenes dataset that inference each scene once, NuPrompt dataset requires testing each prompt, which is time-consuming.
To save time, please download the preprocessed val file nuprompt_infos_val.pkl
.
It contains two random prompts for each scene in the validation set, and the corresponding generation script can be found in ./tools/create_nuprompt_infos_val.py
.
Besides, the other file instance_token_to_id_map.pkl
should be downloaded for easy validation.
After doing so, the structure of that directory will be:
- nuscenes
- v1.0-mini
- v1.0-test
- v1.0-trainval
...
- tracking_forecasting_infos_test.pkl
- tracking_forecasting_infos_train.pkl
- tracking_forecasting_infos_val.pkl
- tracking_forecasting-mini_infos_train.pkl
- tracking_forecasting-mini_infos_val.pkl
...
- nuprompt_v1.0
- nuprompt_infos_val.pkl
- instance_token_to_id_map.pkl
5. Download Single-frame Detection Model [Here]
After downloading, please extract them to ./ckpts/
of the root directory of this repository. The single-frame detectors is provided from the first stage of training PF-Track. Downloading them will save the effort in training single-frame detectors for reproducing our results.