Seems to be an intermittent bug that I come across from time to time - the gmask tracer fails to take a ref frame. Press F8 and it’s blank. I can’t figure out why. Anyone else seen this?
When it happens, usually my axis is selected rather than my gmask node.
Ok. I’ll give it a nudge. Ta.
Tried it - still doesn’t work! It’s wrecking my head. I’m having to go round the houses to track what seems really simple. Have tried region on different values. Multi feature. Perspective grid. And Mocha.
What does work is a 4 point track but I can’t for the life of me get planar to work.
Any advice would be greatly received.
ARS_sh013_fcomp_v020.zip (275.5 KB)
And here’s a jpg sequence of the stuff I’m trying to track in case anyone is bored and wanting to show me how to do it with a planar track.
1 Like
@johnt - The name says it all! 
What? Ref frame blank?
I’m encountering the same problem with your test sequence, yet it works perfectly fine with another image.
I’m starting to think that the stark nature of this grid confuses the algorithm. Creates a corner case where the algorithm tires to judge quality of the reference frame, and throws up its hands.
Have you had the same problem with other images?
Only this one on this scale.
Tried switching Tracking Algorithm to “Multi-Feature Detection?” That sometimes makes it work when it’s refusing to track normally for me.
Yep. Already tried that. Different number of features etc etc.
1 Like
I see this from time to time and it’s so perplexing and frustrating. If it can’t get the ref frame on the first try I go straight to Mocha in Batch.
I think this image does not work with the planar tracker (including multi-feature, tried that also).
Not clear in the end what you do once you have a track. Do you need a planar track?
If a 2-point track is sufficient, track the axis with two points.
If you need more, use Action, perspective surface, which gives you a four-point track. I just tried that on your image, and as expected works perfectly fine.
Also Mocha Pro if you really need a planar track. Though perspective is equivalent as long as you have easy to track features, which you do.
1 Like
Jan, you’re saying that perspective track in Action is the same as Mocha Pro? I did not know that. I don’t think I’ve used perspective since planar was introduced.
Already tried Mocha. Didn’t work either.
Good old 4 point track with some manual tickling did the trick. I’m just curious why something that looks so simple can not be tracked by a planar.
Well - should clarify that. There are two distinguishing features between them.
What transform can you execute, and what features are you tracking.
Perspective tracker and Mocha Pro have equivalent capabilities, but use different ways to get there.
1 point track: translate
2 point track: translate, rotate, scale (in 1 axis)
4 point track: translate, rotate, scale (2 axis), skew
Planar trackers are equivalent to a 4 point track as they create the same transform matrix.
But for a 4 point track (perspective) to work, you need 4 high quality features. Whereas a planar tracker auto-detects multiple features and automatically manages feature lifetime.
A lot of folks default to a planar tracker for its simplicity. You don’t have to think much about it. If it does a good job of course that is.
A four point tracker like perspective is a lot more work, and has a difficult time with some material. But when it does work, gives you a lot more control.
Might be easier to explain with the Silhouette tracker… Each tracker will have it’s own process, but here you can see a bit more of the workings…
1st image: Si Planar tracker in auto mode for perspective tracking.
Min features 200, min distance 3, max age 10 frames.
It finds 433 features. That’s more than the minimum threshold. And it starts tracking fine.
2nd image: tell it to only use corners, only finds 149 features, below the minimum.
3rd image: tell it to only use edges, finds 199 features, just below minimum.
This image has a few challenges.
As it is pretty small relative to the frame, feature points are close to each other, and trackers try to enforce minimum distance to get better math. The further apart the tracking points the more precise the angles get.
Secondly this image is black and white. Most trackers use luma values as input. But there is no variation in the white, it’s all maxed out. So the only things the tracker can hang onto are the edges and corners. That’s not great. If your graphic had some bevels, it might actually work better, or something that adds trackable detail inside the white surface area.
The last thing some planar trackers do is they put a max age on each point, to avoid bad points having a big impact. So once the feature has run for n frames, it needs to get replaced with a new point. But if there aren’t enough good new points to replace it, you run out of features.
And as you see the tracker wants a minimum number of quality features at at any given time.
These are the quality metrics the tracking algorithm will apply to your reference frame. If they don’t come together, you end up with ‘black reference frame’ and it will refuse to track. Many other trackers don’t expose these details, so it’s harder to follow why something falls apart.
That’s where a manual four point track on an image like this works better, because you supervise each one of these points.
Si’s planar tracker worked for 30 frames and then ran out of steam…
1 Like
@johnt - The name of the zip of the clip is ‘ARS’ - if only it meant ART and not BUM.
My apologies for being juvenile.
The only way this will track is to reduce the resolution - try 50% or 25% and then use single points for the corner extents of the matte.
Save the Stabilizer/Tracker data then import it with scale into a new tool at full resolution.
Attach the trackers to the vertices of a bilinear and use UV to nudge something into position (or vice versa if you prefer to work that way).
Export a UV map so you never have to look at such a clip ever again.
As @allklier said the clip is too ambiguous due to its monochromatic nature.
Reducing the clip dimensions through transformation and filtering eliminates much of that ambiguity, permitting the tracker to function without too many false positives.
1 Like
Flames planar tracker worked about the same number before it gave up too.