Modification of the tracker to match our setup

[Renaud Bastien]

Hi,

So we have been trying to see how we could solve our problem of tracking. We have been going deeper into the code trying to understand how the image analysis was done. You gave us the right direction,

https://groups.google.com/g/multicams/c/bwTif3NykIM

but it took some time to understand how everything was fitting together. Right now, we try to understand every step of the image analysis pipeline. That would provide with a good way to identify direction of improvements (either by getting different pictures or by modifying the current algorithm). We are still working on a few obscure bits but we have a few questions that would greatly help us.

First, we have seen many reference to background subtraction. However we have been unable to identify where this function is used and defined in the code. Also the image is thresholded but we could not find where the threshold is defined.

Can you help us?

Best,

renaud

[Andrew Straw]

Hi, Good you are making progress. The actual subtraction (in "absolute difference mode") is done here: https://github.com/strawlab/strand-braid/blob/13b9a5561d85fc99747f6cebf03b1a3af50ef5ac/image-tracker/src/lib.rs#L252-L258 (for "dark" and "light" only modes, see the lines above). The thresholding would be right below here. Again, building on this code (which depends fairly intricately on Intel IPP) is not necessarily the direction I would recommend.

Rather, the more recent "imops" based approach which I have started on is much cleaner. The entire high-level detection code is 30 lines long: https://github.com/strawlab/strand-braid/blob/13b9a5561d85fc99747f6cebf03b1a3af50ef5ac/strand-cam/src/strand-cam.rs#L1204-L1235 and is competitive in speed with IPP (see benchmarks in the imops crate). This isn't (yet) sent to the normal 3D tracking code, but I think that would be fairly straightforward.

Best,

Andrew