We explain the inference procedure on the validation set as examples.
CONFIG_PATHwill be your desired model configurations in configs. It specifies the model architecture.CHECK_POINTwill be the checkpoint of a trained model.PATH_TO_SAVEis the directory to save the 3D MOT inference results and evaluation metrics.
python tools/test_tracking.py CONFIG_PATH CHECK_POINT --jsonfile_prefix PATH_TO_SAVE --eval bbox For example, we can evaluate a PF-Track model designed for ./work_dir/f3_petr_800x320/final.pth, intended to save in ./work_dir/f3_petr_800x320/results/ via the following commands.
python tools/test_tracking.py projects/conf
igs/tracking/petr/f3_q500_800x320.py ./work_dir/f3_petr_800x320/final.pth --jsonfile_prefix ./work_dir/f3_petr_800x320/results --eval bboxYou can use the checkpoint provided by us for a quick try.
- Multi-GPU inference? 3D MOT requires running sequentially on all the frames of nuScenes. Therefore, supporting distributed inference is not straightforward and we does not concern it currently.
- Configuration files?
- Pay attention to the fields of
test_trackingfields in configuration files. - During the inference time, pay attention to
runtime_trackerin the configurations.score_thresholdcontrols the minimum detection score for output,record_thresholdis the score threshold for using track extension, andmax_age_since_updateis both the length for track extension and maximum age for a track before termination.
- Pay attention to the fields of
- Core code? The core functions of tracking happens in [code link], the function
forward_tracking. If you have any difficulty understanding my implementation, please read "my designs," from which you will learn about how to build an end-to-end tracking system.