Using Zero-Weighted Trackers as Canaries

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

‌Using Zero-Weighted Trackers as Canaries ‌

One of the dangers of using complex lens distortion models is that if the model is allowed to become more complex, the best solver solution can explain the data it is given but that is worse at predicting data that it is not given. This is known as overfitting.

You can monitor whether this is happening by using zero-weighted trackers (ZWTs) as canaries (in the way that miners used to use live canaries in a mine to give an early warning for toxic gases).

To do this, select a small collection of trackers to be the canaries, something like 10% of the total, ie 6-10 for a typical auto-tracked scene. If a scene is exclusively supervised-tracked and doesn't have very many trackers, there may not be enough for this technique: you should have at least three.

Change the trackers to ZWTs on the solver panel, and refine the scene (perhaps with some temporary refine fuzz). The ZWTs no longer influence the solve. You'll see the camera/object's overall hpix error, and the solver output will also contain an overall error for the ZWTs.

Now you're ready to try various alternative lens model settings. As you enable more complex lens distortions, you'll typically see the camera/object's overall hpix error go down.

You should also monitor the overall ZWT error, the canary output. Ideally, the canary error decreases also, because the ZWTs can be predicted better also.

If you see the canary error increase, that indicates overfitting. The model is unjustifiably complex for the available tracking data you are giving it, and the solve is overfitting that data. If that happens, you should go back to a simpler lens model, or add substantially more well-distributed trackers, so that it is possible to accurately determine and distinguish among the lens model parameters.

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