search

quantumjot / btrack

Bayesian multi-object tracking
https://btrack.readthedocs.io
MIT License
310 stars 50 forks source link

[BUG] not all objects passed to btrack are tracked - some seem to get lost #405

Open lupinthief opened 2 months ago

lupinthief commented 2 months ago

I pass btrack a list of 2917 objects, each with attributes ['t', 'z', 'x', 'y', 'area', 'umap0', 'umap1', 'xs', 'ys']. I don't run the optimisation and am tracking solely on visual features - here 'area, 'umap0' and 'umap1' (no motion).

The returned tracks contain only 2668 non-dummy points (3990 including dummies). Any idea what has happened to the 249 missing objects that I passed in? I get plenty of tracks of length 1 timestep, so don't think it is missing those. I've buffered the volume considerably so it shouldn't be that any are ever out-of-volume. I can't think what else to try.

I attach the output .h5 file

Your setup: btrack_version: 0.6.4 system_platform: Windows-10-10.0.22621-SP0 system_python: 3.9.18

trackingh5.zip

lupinthief commented 2 months ago

Interestingly, if the motion model is turned on, btrack seems to pick up all the objects passed to it, but not if just using only visual features. When using only visual features, increasing the search radius increases the number of objects that are tracked, but not as far as to pick them all up.