I have some shots that have plenty of information to track. After i use the camera analysis i get what looks on the surface like a perfect track. Although the track looks good on closer inspection there is tiny wobble on the track on quite a few frames and the shot has been kicked back and so it should.
I used the tracker settings of HIGH and set the RES to 1. Is this too much? Can anyone recommend any different settings that would work better to use? Does anyone have any must doâs that will help me other than using syntheyes which is a brain fryer. Iâve got about 60 of these to get through so i really want to do all these in Flame. It would be great to know if anyone has any golden tips to share. Many thanks as always.
Camera tracks are tricky. Sometimes, very rarely, you can average out or delete the wobbly key frames, but most of the time thatâll cause chaos.
Sometimes deleting a ref point or two, or adding a few manually will smooth out tracks, but the best answer is âSyntheyes.â On the bright side, Syntheyes is well priced and fast as hell, so youâre likely to get your time back and then some after you learn itâs tricks.
Hi Andy
Thanks for the tips. I actually own Syntheyes i just need to invest the time in learning it more than just hitting the AUTO button then closing my eyes and praying.
My Syntheyes suggestion is to learn the âsupervised trackersâ which are like flame 2d trackers. Theyâre great and you can use them either standalone or as a supplement to an auto track. Thatâs just about the only part of the program I know and I have pulled some miraculous tracks with them.
I havenât used the new tracker in flame yet so maybe this is irrelevant, but if an auto track isnât working the main thing Iâll try is cranking up the number of points tracked to the max and auto delete points from there based on track error, etc. Also masking out whatever parts of the plate are going to mess up your track like reflections, anything moving. If you have a close enough track and thereâs a little bump donât be afraid to nudge your object a couple of keyframes as well. Some of the other things are undistorting your plate before track + comp as well as finding out your focal length, sensor size, etc. and entering the information to the node.
If you are just yeeting the shot through the tracker and hoping for the best, the best thing you can do is to make sure youâre feeding the solver as high quality 2d tracks as possible. After that first track and solve, go back and delete any tracker that is not on a world-fixed, non-deforming object. That means especially no false intersections (think where something in the fg overlaps a contrasting object in the bkg, like where the frame of a window intersects someoneâs head 10â in front of it), no specular highlights, nothing in reflections, etc. Doing that generally gets me 95% of the way there. HTH!
But once you get the manual tracking under your fingers in syntheyes (I generally auto track, kill bad trackers and then strategically place manual trackers), youâll never go back. So fast. So good. So wow.
If youâre already getting what is mostly a good track with a few bits of wobble then its likely you can get a rock solid track from Flame with some simple adjustments. Its hard to say exactly what those adjustments might need to be without seeing the footage, but one simple change I would recommend when using camera analysis is keeping the default settings.
I did some experiments over the course of several weeks last year with the camera analysis node and concluded that for most situations Fast and Res 2 resulted in a higher quality track than High and Res 1. I also would not use the built in feature masking at all and instead opt for gmasks to restrict the areas to be tracked. It is almost always necessary to adjust the tracker threshold and refine the analysis after the initial track. I usually push the threshold quite hard(sub .1) and the number of trackers that falloff as you adjust can give a good indication of whether the track is worth pursuing further. It may be necessary to manually remove points in order to refine the quality of the track. This can be a painful process with the camera analysis node as the processes of adjusting the threshold and manually disabling points can interfere with each other. I take a more is more approach and disable large swaths of points instead of trying to be selective. It wonât necessarily give a better result than picking through one by one but it also isnât time prohibitive.
Donât trust the colored box indicating the âqualityâ of your result, it is a nearly meangingless indicator(and frankly a bad piece of design). Iâve had bright green tracks that donât work at all and completely red tracks that work a charm.
I would also say that camera analysis is just more prone to âwobbleâ than the mono analyzer. For shots like yours with plenty of information to track, youâre probably better off using the mono analyzer.
Most of the time itâs not as much about the initial parameters, but actually finding and weeding out the points that didnât work. Usually you want an overall deviation value of less than 0.4 for a solid track. Flame and the other tools have ways of filtering tracking points that are above that.
The other thing to look at is refining your camera parameters and lens distortion. They have a huge impact on the track quality in the end. While these trackers all have an âunknownâ mode, and even if you have values from the field, theyâre probably only close estimates and not correct to fractional values. Not sure if Flame has the equivalent of 3DEs attribute refinement workflow?
If you have to do this a lot, itâs worth learning not just Flame but also one of the dedicated trackers. It will take time, but is time well spent and will result in you keeping a bit more hair.
Iâm a big fan of the camera analysis but its a scalpel to the mono analyzerâs swiss army knife. I tend to think of it more as a texture tracker or an object tracker rather than a camera tracker.
This is definitely a laymanâs intuition, but I feel that the reason a red(bad) track can work well and a green(good) track can fail at times is that the camera has a tendency to solve for a particular cluster of well tracked points around a single feature, usually the most dominant and consistent texture, throughout the shot. Sometimes this solve is also accurate to the outlier points and sometimes it isnât. This can cause what is a good track for, hopefully, the feature you intend to project upon to appear bad(red) because it doesnât solve for any of the points outside of the feature. Likewise, although less common, a solve can be accurate for most points and appear to be good(green), but be a bad solve for the specific points you are trying to project onto.
I almost always try both camera trackers when I need to track a shot. There will always be times where one node outperforms the other in ways that are inexplicable. That said I believe that the mono analyzer is still better in the aggregate. The camera analysis node is better for tracking shots where one point tracking is compromised due to focus, camera shake, object occlusion, etc. Also while the mono analyzer can usually be massaged and is at least useful if you can get the error sub 2, the camera analysis is more of an all or nothing kind of deal. Iâve seen it align a plane along the surface of a building on par with what you might get back from an outsourced track and other times not so much.
I have a DJI Mini 3 Pro showing up tomorrow( ) so maybe Iâll take it for a spin this weekend and upload a video of how I approach the trackers. The Autodesk tutorials are useful for learning where the buttons are, but they only ever track shots which donât need a camera track so not particularly helpful for learning how the nodes actually behave.
This would be amazing. Im constantly trying to get both trackers to work well. Sometimes shots that i think they would smash end up awful and vice versa
@Marcus_M Aside from the amazing Logik Live episodes with the most relevant info for us in Flame, Iâm finding this guyâs YouTube channel pretty good and filled with small but really helpful tips: https://www.youtube.com/@MatthewMerkovich
Iâm so going to watch these vids thanks all very much. Tracking is such a dark art so sharing knowledge is ace. So it turned out after reading Rylands 1st email that i should dust off the mono analyzer node. I fell victim to ânew must be betterâ with the camera analysis tracker and i was very wrong. After using the mono analyzer, deleting bad points refining to an error of 1.3 ove got the track i need. Infact a bang on track. I defo now will keep an open mind on what tracker i should use. Thnk you for all the tips and the links. Keep the tracking tips coming.
) so maybe Iâll take it for a spin this weekend and upload a video of how I approach the trackers. The Autodesk tutorials are useful for learning where the buttons are, but they only ever track shots which donât need a camera track so not particularly helpful for learning how the nodes actually behave.