STIRMetrics is an evaluation framework for the STIR Challenge accuracy metrics to evaluate point tracking in surgical scenarios. STIRMetrics provides baselines to get started. For 2D, we provide implementations of CSRT, MFT, and RAFT. For 3D, we provide a simple RAFT+RAFT-Stereo method. Modify the code as needed to run your tracking algorithm and output point tracks.
See the STIR Challenge for details.
Install STIRLoader using pip. If using MFT, install the MFT adapter MFT_STIR, or if you would like to run models without MFT, comment out all usages of MFT from the codebase.
Configuration: Edit config.json to give the STIR dataset directory and the output directory.
Set datadir to point to the extracted STIR validation dataset directory.
Set outputdir to point to a folder in which the output estimates will be written.
git clone STIRMetrics
cd STIRMetrics/src
All commands can be run from the src directory.
Follow the docker installation instructions at STIRHoloscan. Afterwards, you can use the rundocker script here:
./rundocker.sh <path_to_data> <path_to_outputdir>
cd src
Now you should be able to run the same commands (with visualization disabled via passing the -showvis 0 flag to the python applications).
A click-and track application for visualizing a tracker on STIR. Example usage to track on three clips:
python datatest/clicktracks.py --num_data 3
Uses the labelled STIR segments to evaluate tracking performance of given model.
python datatest/flow2d.py --num_data 4 --showvis 1
This produces an output json of point locations to the output directory.
An extension of the flow2d that evaluates the tracking performance in 3D as well.
python datatest/flow3d.py --num_data 4 --showvis 1
This produces output json files of point locations in 2D and 3D (in mm) to the output directory.
Firstly, generate ground truth json locations using the following command. Swap write2d for write3d to generate the 3D positions. Files are written to the ./results directory.
python datatest/write2dgtjson.py --num_data -1 --jsonsuffix test
For challenge submissions on the test set, the ground truth location files will be generated, and run independently by the challenge organizers.
Then generate a json file of tracked points predicted by your model using (select whatever you want when testing for num_data, and modify the random seed for different sets).
python datatest/flow2d.py --num_data 7 --showvis 0 --jsonsuffix test --modeltype <YOURMODEL> --ontestingset 1
or for 3d:
python datatest/flow3d.py --num_data 7 --showvis 0 --jsonsuffix test --modeltype <YOURMODEL> --ontestingset 1
For the challenge, we will run your model on the testing set with --num_data -1. Ensure your model executes in a reasonable amount of time. Set your modeltype, or use MFT, RAFT, CSRT to see baseline results for 2D (RAFT_Stereo_RAFT is the only modeltype baseline available for 3D). ontestingset must be set to 1 for the docker submission, since your model will not have access to the ending segmentation locations. For the challenge we will be running your model via the flow2d/3d commands. Thus we recommend not modifying the command-line interface to this file.
The generated ground truth files (start and end gt locations) and your estimates can then be passed into the metric calculator with this:
python datatest/calculate_error_from_json2d.py --startgt results/gt_positions_start_all_test.json --endgt results/gt_positions_end_all_test.json --model_predictions results/positions_<numdata><YOURMODEL>test.json
python datatest/calculate_error_from_json3d.py --startgt results/gt_3d_positions_start_all_test.json --endgt results/gt_3d_positions_end_all_test.json --model_predictions results/positions3d_<numdata><YOURMODEL>test.json
This will print threshold metrics for your model, alongside metrics for control version of zero-motion. For the challenge, this script will be run by organizers on your positions.json file. For the 3D submissions, we will evaluate metrics for both the 2d and 3d locations.