< Previous | Contents | Manuals Home | Boris FX | Next >
You can select a neural network before starting auto-tracking, from selectors on the Summary Panel (at the top, below the motion-type selector), on the Features panel (again at the top), and on the Advanced Feature dialog (launched from the Features panel, further down after the tracker count controls).
Important : The pixel aspect ratio should be 1.0 for neural-net processing. Use the image preprocessor’s Output tab to set up resampling to make that the case.
Once you’ve selected a neural net, you can start auto-tracking by the usual AUTO button, Run Autotracker, Blips playback range, or Blips all frames, and the neural net will be used instead of the normal (fast!) spot-based autotracker.
Warning : Auto-tracking currently requires that neural net results be generated for every pixel of the input image. That means it can be much slower than the traditional auto-tracker. (This isn’t the case for supervised tracking.)
On the Summary and Features panels, there’s only the neural net model selector. But on the Advanced Feature dialog, once you’ve selected a model, the non- neural controls are removed and replaced by a set of neural net controls.
The primary neural-net-specific controls are the Blip Density, which indicates (limits) how many will be generated in an area (as for regular auto-tracking), and the Minimum amplitude, which controls the sensitivity: how “blippy” a pixel should be to be usable as a blip. You can turn on the Advanced Feature dialog’s Auto Re-blip checkbox and try adjusting the controls, though some patience will be required.
The tiling controls indirectly affect performance, not the result, but also affect the amount of memory required.
Tip : If autotracking runs very quickly and produces no tracks, you must adjust the Tiling controls, typically the Tiles per Batch!
Processing a neural network requires temporary storage many (hundreds?!) times larger than the size of the image being processed. If too much memory would be required, the neural net processing will fail outright and produce no result. To avoid that, we break the images into tiles, and a batch of tiles are presented to the neural net for processing.
However, the smaller the tiles, the more extra work that must be done, processing an overlapping margin between tiles. If the tiles are too big, even if they don’t run out of memory, they can create inefficiency along the right and bottom edges of the image, as the tile size doesn’t usually exactly divide the image size (minus margins on all four sides).
The ”best” choices can’t be determined in advance, they depend on the details of your image, computer, and video cards. Fortunately as a practical matter the results typically don’t depend all that strongly on the settings.
The most important thing is that if the neural net is failing, quickly producing no results (and maybe an error message flashing by in the status bar), you need to reduce the Tiles per Batch. In rare cases, adjusting the tile width by one or two may matter also.
Note that the controls may initially be set to values determined from preferences, and they are reset to the neural net’s default values each time you change the neural net selection (whether the Advanced Feature Control panel is open or not).
©2024 Boris FX, Inc. — UNOFFICIAL — Converted from original PDF.