Changelog

20.7.0

  • Correct the use of arccos function to be compatible with numpy<=2.0.0.
  • Test with numpy<=2.0.0.

20.6.0

  • Define, debug and optimize the implementation of IoU-like-two-yaw metric.
  • Make the new experimental metric (IoU-like-two-yaw) available.
  • Remove the dangling average dim IoU metric.
  • Add argument ind_pos_w_rz_size to customize the GIoU-yaw and IoU-like two yaw metrics.
  • One less temporary array in the real IoU yaw metric.

20.5.0 (yanked)

  • Add an experimental IoU-like two yaw association metric.
  • Update (development) dependencies.

20.4.1

  • Define range of Python versions >=3.11.
  • Remove slow code for GIoU, and GIoU-like.

20.3.0

  • Using vectorized convex hulls from shapely.
  • Using shapely.area u function.
  • Using shapely.multipoints u function.
  • Using shapely.polygons u function.
  • Vectorizing the intersection and union in 1D.
  • Do not repeat computation of the first volume.
  • Added F3T typing alias.

20.2.0

  • Start to exploit the vectorized capabilities of shapely.

20.1.0

  • Bugfix: ensure the number of detections can be zero.

20.0.0

  • Add yaw-dependent GIoU 3D metric giou-yaw.
  • Rename size-modulated-mahalanobis to mahalanobis-size-modulated.
  • Rename giou -> giou-axes-aligned.
  • Remove the unknown metrics and matching algorithms.
  • Remove mean-dim GIoU metric.

19.3.0

  • Check the annotation indices for type, dimensions, data type and size.
  • Return deepcopy of the ScoreCopy objects.
  • Return StubScoreCopy if scores are not needed (faster).
  • Measurement covariance is allocated at instantiation. NaN on diagonal to enforce definition.
  • Possibility of association within given object classes.
  • Refactor the ScoreCopy class for multiple classes.
  • Demand return_score flag at tracker instantiation.
  • Remove return_score flag from the keyword arguments of the .advance method.
  • Remove .tracks (list of tracks) from tracker (breaking change for most consumers).
  • Add array of lists of tracks as .tracks_c attribute.

19.0.0, 19.1.0, 19.2.0

  • Yanked as not completely usable.

18.0.4

  • Include OpenCV via the mechanism of optional dependencies (extras).
  • Adjust the dependencies to tolerate “opencv-python>=4.11.0.86”.

18.0.3

  • Adjust the pyproject.toml to explicitly tolerate “numpy>=1.22”.

18.0.2

  • Update opencv-python-headless dependency since it is not linking against X11 anymore.

18.0.1

  • Declare opencv-python-headless as dependency, not opencv-python.
  • Add “Use the kinematic-tracker as dependency” section in the README.

17.0.1

  • Optimize the dependencies: relax numpy version constrains.

17.0.0

  • The advance method returns the score (metric) if the keyword argument return_score=True.
  • The advance method returns either ScoreCopy object or None.
  • The advance method does not return the list of tracks. Tracks are always available via the attribute.
  • Update the dependencies.

16.1.0

  • Reset method in the tracker.
  • Shortcut for importing the tracker class from kinematic_tracker import NdKkfTracker.
  • Update the dependencies.

16.0.0

  • Assert the number of variables in the measurement vector is greater than 5 in GIoU aligned.
  • Change the default association metric to be Mahalanobis (previously GIoU).
  • Link to the documentation on GitHub Pages.
  • Make the annotation IDs optional.
  • Improve typing hints in the method NdKkfTracker.advance().

15.7.1

  • Improve the definition of the dependencies to avoid compilation on Python 3.14

15.7.0

  • Introduce generator of measurement vectors in MetricDriverBase class.
  • Precompute the GIoU auxiliaries in aligned GIoU driver.
  • Setter for the max number of reports and targets.
  • Suggest a dimension-averaged GIoU association metric.
  • Improve type annotations in metric drivers.
  • Improve testing of the current definition of the size-modulated Mahalanobis.

15.4.0

  • Suggest a size-modulated Mahalanobis metric.
  • Improve the computation of the regular Mahalanobis metric.
  • Relax requirements for fancy indexes in the tracker.
  • Add possibility to set the fancy indices and metric type at once.
  • Assert that empty detections are acceptable for all metric drivers.

15.0.1

  • Using the package association-quality-clavia.
  • Published in PyPI registry.

14.2.0

  • Using the package binary-classification-ratios.

14.1.2

  • Compatibility with nuScenes @ evaluation-against-nuscenes (Python >=3.10 & NumPy >=1.26.4).

14.1.0

  • Added an association-threshold setter in NdKkfTracker.

14.0.1

  • Corrected the example of tracking with yaw angle affected by the version 14.
  • Made the example of tracking association quality more robust.

14.0.0

This version bump is major because it breaks a silent convention on the distribution of variables in the measurement vector. In previous versions, the measurement vector has been distributed as follows:

 $$z^T = (p_x, p_y, p_z, s_x, s_y, s_z, ...)$$

In this version, the distribution of variables in the measurement vector is as follows:

 $$z^T = (p_x, p_y, p_z, ..., s_x, s_y, s_z)$$

where the position variables are listed first, followed by the size variables. This change makes the tracker more convenient to use with ASAM OpenLabel frames.

  • Possibility to adjust the distribution of variables in the measurement vector.
  • Added setter NdKkfTracker.set_ind_pos_size.

13.0.1

  • Proofread the README file.
  • Renamed gen_z –> gen_xz in tracker.

12.1.0

  • Remove the unfinished GIoUWithYaw class.
  • Add all non-trivial docstrings.
  • Eliminate the flip in Greedy matching.
  • Eliminate unused buffer in the Mahalanobis metric driver.
  • Remove the obsolete method NdKkfTrack.get_det_cov().

12.0.0

  • All dependencies are public from PyPI.

11.1.2

  • Add docstrings in nd modules
  • Use documented version of the kinematic-matrix package (viulib-kinematic-matrices>=0.3.0)

11.1.0

  • Allow both int and np.int64 to be used as time stamps in the Tracker.advance method.
  • Add three tutorial examples:
    • tracking of cuboids aligned with Cartesian axes
    • tracking of cuboids with yaw angle
    • tracking and classifying the association quality

11.0.0

  • Introduced the (bipartite) matching drivers:
    • Maximal shape of the metric is rectangular
    • The returned match is in fixed-size numpy arrays which stay the same across time stamps
    • Auxiliary buffers are used instead of temporary (scrap) RAM
  • Remove the calculation of threshold from the association probability

10.4.0

  • Added nuScenes example
  • Refactored the metric drivers:
    • Maximal shape of the metric is rectangular
    • Metric is computed in a chunk of memory at the beginning of the rectangular buffer
    • The returned matrix is rectangular, local and shares memory with the buffer
    • Removed O(N) buffers for $Hx$ and $HPH^T$
    • The order of indices is (reports, targets) resulting in wide matrices most of the time
    • Auxiliary buffers are used instead of temporary (scrap) RAM

10.3.0

  • Add ClassificationRatios helper class
  • Add fusion-lab example of two almost colliding cuboids
  • Add padding to the Hungarian matching
  • Association-quality- and Cartesian-quality helper classes in unit tests

10.2.0

  • Add AssociationQuality classifier
  • Change default association threshold from 0.9 –> 0.25
  • Fix finite-diff adaptive noise to multiply with the variance factor
  • Add setter NdKkfTracker.set_measurement_std_dev(…)

10.0.0

  • Removed GatherAtStart in favor of DerivativeSorted

2.0.2

  • Started GIoU with yaw

1.1.2

  • Initial release
This site uses Just the Docs, a documentation theme for Jekyll.