Open 9a3eedi opened 8 years ago
Dr. Tarek mentioned that this is probably because we're using SSD for tracking, which needs to look at every single picture in the image, which can be slow. Using another algorithm, or tweaking the existing algorithm (so that it downsamples for example) might be a way to fix this problem.
I rewrote the code so that it's easy to switch between the different tracking algorithms. I put the following tracking algorithms provided by VISP:
A quick eyeball test seems to show that SSD Inverse Compositional gives the best speed and performance, but a more formal test is needed, so I'll put some timing code and try to generate a report.
When the template size is large (i.e. when the detector detects a marker that spans a large percetnage of the image), the tracker becomes very slow, with it being only able to track around 3 times per second. This needs to be improved.