Skip to content

Latest commit

 

History

History
28 lines (21 loc) · 1.99 KB

inference.md

File metadata and controls

28 lines (21 loc) · 1.99 KB

Inference

We explain the inference procedure on the validation set as examples.

1. Inference Commands.

  • CONFIG_PATH will be your desired model configurations in configs. It specifies the model architecture.
  • CHECK_POINT will be the checkpoint of a trained model.
  • PATH_TO_SAVE is 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 $800\times 320$ resolution (small-resolution setting), trained at location ./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 bbox

You can use the checkpoint provided by us for a quick try.

2. Notes for Developers.

  • 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_tracking fields in configuration files.
    • During the inference time, pay attention to runtime_tracker in the configurations. score_threshold controls the minimum detection score for output, record_threshold is the score threshold for using track extension, and max_age_since_update is both the length for track extension and maximum age for a track before termination.
  • 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.