search

dstl / Stone-Soup

A software project to provide the target tracking community with a framework for the development and testing of tracking algorithms.
https://stonesoup.rtfd.io
MIT License
384 stars 125 forks source link

Low efficiency in GOSPA metric computations #1032

Closed narykov closed 16 hours ago

narykov commented 3 weeks ago

I'm using the GOSPA metric to evaluate the tracking results in a scenario that has generated 40 tracks over 30,000 time steps, and I've been waiting for results for more than a couple of hours. I understand that this isn't a trivial scenario, but it made me think about how efficient is the metric's implementation.

The description indicates that the metric relies on the auction algorithm, which may not be the fastest option here. Angel Garcia-Fernandez provides an improved (MATLAB) implementation that uses the assign2D function instead. It would be great to have it in Stone Soup, provided that assign2D has already been implemented to speed up the GNN algorithm.

sdhiscocks commented 3 weeks ago

I've opened #1033 which switches out the algorithm for the SciPy implementation (which is based off of linked MATLAB version).

Just as another suggestion, you could reduced/interpolated truth and tracks to reduce the number of timestamps to speed up metrics, if you don't really need metric at every timestep.