How can you "fill in" untrackable keyframes in PlanarTracker

I have a situation where a camera's in motion and the plane I need put a corner-pin on is untrackable for the first third of the shot. But, for that section, a simple translation/rotation/scale track will suffice. But I can't find any way to "fill in" tracking data for the problematic section. If I go to the frame before which I need to prepend tracking data and try to track backward with the PlanarTracker node using a different method, you can't do that because it says it will delete all tracking data.

So... what to do here? I can't just set up a separate tracking node with keyframes only populated during the problematic range, because the sizing and warping of the corner-pinned image is done by the initial PlanarTracker.

So I need to take the image that's shrunken down by the corner-pin and track it into place with a simple track before the PlanarTracker's motion keyframes kick in.

Suggestions?

Here's the shot. After this I can CornerPin. Before it, I can just do a primitive track using the edge of the computer.

trackingProb.jpg (279.28 KiB) Viewed 1000 times

The Corner Pin can be adjusted using the Reference Time Position tab. So if you just track something that is as close to moving in a similar motion or the same plane, to get the tracking keyframes, then you should be able to keyframe the Corner Pin into position, using the Reference Time Position controls. That's if I understood your problem correctly.
Or you could use the Corner Pin in the Point Tracker but I don't think the Corner Pin in one is any more difficult to adjust than the other.

Use the Magic Mask tool, or something similar, to select the monitor and its features that can be tracked. Then, track only those selected parts. This turns the monitor into its own occlusion mask. This is simpler than creating separate masks to block each thing that might be behind the monitor.

Thanks for the replies. I tried hand-correcting the cornerpin during the untrackable region, but the image visibly drifts on the screen. That's why all I really need to do is take the shrunken-down image and attach it to a plain two-point tracker that is tied to the corners of the laptop screen. But I need to acquire that image from the existing PlanarTracker, because it's doing the scaling and positioning.

To do this, I even tried copying the PlanarTracker, deleting all keyframes but the first, and then using that node's output to provide the scaled and positioned image to superimpose; I then created a PlanarTracker using points attached to the laptop corners and translation/scale/rotation to track the little image into place during the first part. But this produced wildly incorrect rotation values for the superimposed image; no idea why.

As far as the Magic Mask suggestion goes... I already selected the area that can be tracked, using the PlanarTracker's polygon-based selection area. There just isn't enough surface to track before the point depicted in the screen shot.

Is the shrunken image itself moving footage or a still? If it's a still just use a timestretcher to get the output of the planar tracker at the frame you need it and proceed as you describe. If you're doing the actual merge in the planar tracker now, change that so you only output the Foreground and then to do the merge in a separate merge. This will give you access to the output of the planar track itself. That's all there is to it, really. Your approach to the approach is correct as far as I can tell.

Thanks for reading all that and weighing in!

The image to be superimposed is moving footage. That's why I need to use its scaling and position from the PlanarTracker that's doing the corner-pin for the last 2/3 of the shot.

By copying the PlanarTracker (and its data) I did gain access to the small image, similarly to the strategy you suggested. And by deleting all but its first keyframe, it remained in the correct place to start that last 2/3 of the shot.

But it went off the rails when I tried to use another tracker for the first 1/3 and move the small image with it. It was unclear why the resulting rotation and scaling from a two-point track was so absurdly incorrect, but I finally got fed up with it and just started over in Nuke.