Abnormally high camera solve error

Not sure if this post should go here or not, but it is technically animation, so here it is…

I’m going through the Track, Match, Blend dvd training in an attempt to master blender’s tracker and I’ve noticed with certain footage I sometimes get crazy high solve errors even when all my markers seem to be tracking at least decently well. I’ve had upwards of 2000 before, but more commonly I get 200-600 on these problem shots.
Now, I’m in the middle of the training chapter on feature detection and made sure my settings are the same as the author’s. Before solving, the first thing I notice is that my curve/graph/chart thingie looks a bit more messy than his (before cleaning up and deleting shaky trackers), but after cleanup, I figure it looks about the same. So I solve at the same time and with the exact same settings as he does. He gets a solve error of 17. I get a solve error of 250.
NOW…I lied. I used slightly bigger trackers than he did. I believe he used size 51 while I used 61. This brings us to the question of the ages…does size matter?? and if not, what other factors would contribute to this monstrous difference in solve errors?

I’ve tried various variations to try to get better results (switching from ‘keyframe’ to ‘previous frame’ and back again, hand-placing trackers, using some larger and some smaller trackers, etc), and as it stands now, I’ve currently got a solve error of 148. whoopty flippin doo.

To wrap this up, I have gotten great solves in the past. I’m just having trouble determining all the necessary factors that lead to such a solve more often than not.

Edit O and also I’ve been getting this error lately – ‘some data failed to reconstruct. see console for details’ where the console says ‘no bundle for track #…’ and says this for a few tracks. Not sure if it’s related, but figured I’d mention it.

bump I might start taking it personal if nobody replies…sad face

this appeal to emotion brought to you by kleenex.

I have marker errors in the 10–15 range and the camera solve is 35,000 :spin:

Is your lens length and image area a match for the source? Are there any jumps in the track, sometimes I try to track an element of the image that repeats elsewhere. Like a brick pattern. Poor old tracker finds the same element and skips over to it.

Sensor size and focal length are set according to specs I found on the internet. It’s difficult because it’s a phone camera (albeit a good one). It may be an issue with rolling shutter distortion. It’s also a difficult scene that is perspective only (traveling down a road). I stabilized the 2d footage and cropped the images. I tried moving the optical center because of that, but it got worse.

Optical centre exists only for cropped photos I think, so don’t use that for video. You should not try reconstruction on stabilised footage, as it may be distorted to much for sensible results. If there is bad rolling shutter on a few frames then try manipulating those points by hand or just delete them, letting the tracker average the result there.

Check out the tracker graph display, you should see the offending frames by the deviation on the timeline.

To reduce the Solve Error, did you use a smaller sample of frames to Solve the Camera Motion?

And what about when a tracker goes out of frame and comes back in? That’s my big question.