- Update the tutorial.
-
Clearify the flow on the evaluating frame by frame.
-
Fix several bugs on evaluating metrics.
-
Add a new by-frame function, read_detections_per_frame_v2, on the IOUTracker class. This function improves a lot the matching speed between a tracker and a detection. It upgraded the method purposed by the original paper. It allowed the detection boxes of each frame recorded itself tracker IDs while in applying the IOU tracker algorithm.
-
Move the original caller of the IOUTracker object to the previous method (
__previous__
). Pointed the latest version of IOUTracker with the results annotated with Tracker IDs.
-
Updated the tutorial Jupyter notebook.
-
Add an assertion to make sure that the data type of UID in the ground truth or the prediction must be an integer or a string.
-
Fixed several bugs on the evaluation flow.
- Update the class EvaluateByFrame and add a function evaluateOnPredsWithTrackerID to it. It allows evaluating the tracking result with the ground truth without the IOUTracker algorithm involved.
-
Fix a bug on the GTTrajectory class. Solve the first time adding the detection box, and it causes a IDSW.
-
Fix a bug on the additionApproach on the Hungarian class.
IOU Tracker 1.1.0 improves easy-to-use functionality, including auto tracker ID increment and redesign metric and helper APIs.
-
Move the matching algorithm, Hungarian, from the metrics.MOTmetrics to the src.helpers, and its unit tests as well.
-
Move the BBoxIOU and a static method of the IOUTracker class, detections_transform, to the src.Helpers class.
-
Add the auto tracker ID increment as the default feature.
-
Add an implementation to call IOUTracker for returning the corresponding track information, including ID, total time period on the unit of the frame, and a flag for considering the validation.
-
Update the tutorial.
-
Add more functionalities or APIs to the package ioutracker.
The major algorithm features and metrics were designed and implemented.