‌Suggested Usage

Clearly the neural net stuff is “bleeding edge.” Though neural nets are used commercially, it looks like they are tuned to run on one particular combination of CPU and GPU, and Google, Facebook, and Apple set up specific servers for particular uses. Or, they run in tightly controlled phone environments. That’s a lot different than SynthEyes, which has to run not only on widely disparate machines from laptops to desktops to servers, but on different operating systems as well. We’ve given it a shot anyway!

In most cases, it will be fastest and best to use SynthEyes’s traditional auto- tracker, which is far faster than the neural nets and produces longer and more accurate trackers as well.

The neural net approach may be useful tracking buildings, and other very artificial scenes, when you want to do set reconstruction. The neural nets try to produce features that are salient, ie the actual corners of buildings, windows, and other details. You should see the tutorials on set reconstruction for information on how to do that: you should do initial tracks with small numbers of trackers, then use Add Many to create many more zero-weighted trackers, instead of bogging down with many trackers to start with.

The X-finding net might be worthwhile for green-screen shots with those marks.

You may find it helpful to filter the image, then drop it’s resolution, using the image preprocessor. That can reduce the effect of noise, especially compression noise, in hopes of producing more stable trackers.

Similarly, the using the auto-tracker with the checkerboard-finding net may produce better results on degraded checkerboards than the built-in Shot/Create Lens Grid Trackers detector.

If you have an idea for other common features that a neural net could look for, let us know. Just keep in mind that the limitations discussed above will still apply.

< Previous | Contents | Manuals Home | Boris FX | Next >