Appearance
Fills, seams and holes
This page covers what happens to a gap, what happens at the edges of a repair, and what happens when Post decides not to fill something. A take, throughout, is one recorded motion-capture performance.
Fills
A gap is a stretch of frames where a marker was not seen. A fill is the motion Post writes into that gap, worked out from the data around it. Which step does the filling depends on how long the gap is. That is why the Cleanup Plan, the per-take list of steps Post will take, shows Fill small gaps, Fill medium gaps and Fill long gaps as three separate rows: they are different pieces of work with different prerequisites.
The join matches the speed, not just the position
A fill that only meets the captured frames at the right place still pops, because it arrives at the wrong speed. Post matches the motion at both edges as well as the position, so the repaired trajectory leaves and re-enters the captured data without a step. Where only one side of the gap has data, the fill carries that side's motion into the hole for a short band and then tapers, rather than snapping to a dead hold.
A flickering marker is filled as one stretch
A marker that keeps dropping out and returning used to be filled gap by gap, which left a small step at every boundary. Where the surrounding capture supports it, that whole run is now filled as one continuous stretch. Where it does not, the gaps are still filled separately.
Filling a run as one stretch means a few frames your raw take still shows get replaced along with the gaps. Your raw take is the file you uploaded, which Post never modifies. Those frames are the little islands between the drop-outs. They are reacquisition wobble, and they are a worse thing to anchor a fill on than nothing at all.
Where to see the frames a fill replaced
The Review screen is where you read a finished run. Its region cards say this on the face of the card. A region is one span of frames on one marker that Post worked on.
Includes 111 measured frames replaced during filling (flickering signal)
Expand the card and each stretch is listed with the frames it covers, the check that consumed it, and how far the fill ended up from the samples it replaced:
f962–f1072 · 111 frames · acceleration spike · fill 0.42 cm from measured
One line then calls out the worst of them:
Fill landed up to 1.2 cm from the samples it replaced.
Some deviation is expected, because those samples were consumed precisely for being noisy. The number is there so that a fill which went somewhere the measurement never did is visible.
A card that says nothing about this is a fill that replaced no measured frames.
Sometimes Post first marks one of those stretches for replacement and then keeps it after all, because the diagnosis did not flag it as wrong and the fill can join it smoothly. The card then says the stretch was kept, not replaced, and gives the farthest distance any of its frames moved:
Kapnetix kept 7 captured frames inside the gap within 0.4 cm of where they were recorded and joined the fill to them.
If only some of the stretches were kept, the same note goes on to say how many measured frames the fill still replaced, because the diagnosis marked them as wrong.
Best-attempt fills
Sometimes the usual repair of a gap cannot be completed. Its rebuilt frames may not join the captured frames smoothly, they may move the marker faster than a real body part can move, or they may change too many captured frames next to the gap. Post used to leave such a gap open. It now fills it as a best attempt with a simple curve between the captured frames on either side, or, at the start or end of a take, continues the marker from the nearest captured frame.
A best-attempt fill is a plausible bridge between the captured frames. It is not a reconstruction of the motion that was recorded, so it can look right and still be wrong. Its region card says so:
Filled this 9-frame gap as a best attempt. The usual repair could not be completed here because it would have changed too many captured frames next to the gap. The fill is a plausible bridge between the captured frames rather than a reconstruction of the recorded motion.
The model line under it reads Model · Best-attempt fill, and the card ends with Check these frames by eye before you export. It shows no score, because the words are the warning, and the region counts as a region that needs eyes until you have looked at it. Revert removes the fill and leaves the gap open again.
A gap whose every simple fill would move the marker faster than a real body part can move is still left open, and its card says why.
Gaps the diagnosis did not flag
The diagnosis is the first pass over your take. It finds the problems the cleanup then repairs. Now and then a marker is missing on a few frames that no finding of the diagnosis covers. Post used to leave those frames open. It now attempts them with its usual gap model, and because no finding vouched for the gap, the fill is a best attempt. Its region card says so:
Filled this 6-frame gap as a best attempt. The diagnosis did not flag this gap, so Kapnetix attempted it with its usual gap model.
The card ends with Check these frames by eye before you export., shows no score, and the region counts as a region that needs eyes until you have looked at it. The panel above the region cards gives the take total, for example The diagnosis did not flag 1 gap (6 frames), so Kapnetix filled it with its gap model as a best attempt. Check it by eye before you export.
Stretches where a marker left its path
Beside a spike the cleanup could not smooth, or a gap it could not fill, a marker sometimes leaves its path for a stretch of frames and then comes back. While it is away it follows something else, such as another marker or a stray reflection, instead of the body part it is attached to. Post measures this against the steady markers of the same body part. Where it finds such a stretch, it replaces the whole stretch, or, when the frames belong to another marker of the same performer, moves them to that marker and refills this one. The region card says which, how far the marker had strayed, and how the frames were refilled:
Replaced a stretch of 16 captured frames where this marker had left its path, by up to 21.4 cm, beside a spike the cleanup could not smooth. Those frames did not follow any marker on the performer, so Kapnetix refilled them from the markers on the same body part.
These are captured frames that Post replaced, so the card names them, and Revert puts them back. The model line reads Model · off-path refill. The panel above the region cards gives the take total, for example Kapnetix found 2 stretches of captured frames where a marker had left its path and followed something else. It replaced 16 captured frames in 1 of them and left the other one as recorded because the marker never came back to its path.
Seams
A seam is the join at the edge of a run of frames Post wrote. It is a seam whatever is on the other side of it, whether that is captured data, another repair, or the end of the take.
The Smooth joins step, which is counted in joins, judges each of those boundaries and conditions the ones that need it. The joins it could not land are reported rather than smoothed over. The inside of a repair is checked separately, so every frame Post wrote is judged exactly once.
A join is where a fill is most likely to give itself away. That is why it is its own step and its own check rather than a side effect of filling.
Joins that still change speed too sharply
After every repair has run, Post takes one last smoothing pass over every join. Each join has a limit on how sharply the marker may change speed there: three times the marker's own usual acceleration around the join, and never less than 3 g. This is a limit Kapnetix sets for a smooth join, not a physical limit. Where a join is still over it, the region card says so, for example:
At frame 563, where the fill meets the captured frames, the marker's change in speed is still over the limit Kapnetix sets for a join (3.7 g against a limit of 3.0 g).
Frames are listed from first to last. The figures are given where there is one join. The card then says why the join could not be smoothed, one sentence per reason, and ends with Check this join by eye before you export. A region with such a join counts as a region that needs eyes, so a take that has one is never marked No review needed. The reasons are:
- Every way of smoothing the join would have moved the marker faster than a real body part can move.
- The join meets a short stretch of captured frames inside the gap that Kapnetix kept as recorded.
- Smoothing the join would have made another part of the motion worse, so Kapnetix left it as it is.
- Adjusting only the filled frames could have smoothed the join, but none of the cleanup's smoothing steps made that adjustment.
- The last smoothing pass only adjusts frames that were missing from the capture. The frames next to this join were captured, and this repair replaced them, so they are counted as changed and there is no filled frame left to adjust. This is what a repair that replaced captured frames says, such as a jump that was cut out and refilled. Where the captured frames next to the join are ones Post moved under the gap-edge rule, the card names them instead, for example Next to the gap, Kapnetix had already moved 1 captured frame by at most 0.1 cm to join the fill to the capture and counts it as changed, so there is no filled frame left to adjust at this join.
- Only moving up to 2 captured frames next to the gap, each by a small distance, would smooth the join, and the last smoothing pass does not move captured frames.
- Not even moving the captured frames next to the gap would smooth the join, because the jump is in the captured motion itself.
The panel above the region cards says how many joins the last pass smoothed, and how many still change speed too sharply, each only when the count is not zero. The second count is the joins the region cards name, so it never disagrees with them or with the take's review status.
Captured frames beside a gap
The frame or two right next to a gap are the least reliable frames a marker has, because the cameras are just losing or just finding it. To land a fill smoothly, Post may move up to two frames on each side of a gap, each by at most a short distance. It never does this silently. The region card says it in one sentence, with the distance in centimetres:
Kapnetix moved 2 captured frames next to the gap by at most 0.4 cm so the fill joins the capture smoothly.
A repair that moved edge frames is still a repair, and it does not lower the take's confidence figure. The Review screen gives the take total under the take name, for example To join fills to the capture, Kapnetix moved 4 captured frames next to gaps, the farthest by 0.4 cm. When no frame moved there is no such line, and a card that says nothing about this moved no captured frames.
Holes left open
Post repairs occlusions. It does not invent frames that were never recorded. So a cleaned take can still show holes, and re-cleaning will not fill them. That is the thing worth knowing here, because the instinct is to run it again.
The Review screen names them. A note above the region list reads:
8 windows · 88 frames · 0.7s not repaired
The note then says what each of them is. There are three kinds, and Post tells them apart rather than calling everything the same thing.
| Kind | What it is |
|---|---|
| Out-of-volume window | The performer had left the capture volume. Motion that was never seen cannot be recovered. |
| Capture dropout | The capture system recorded nothing across those frames, so nobody went anywhere. There was nothing to repair from. A window that takes more than one performer down at exactly the same frames is this, not an exit. |
| Not assessed | Detection did not complete on that layer, so nothing there was cleaned either. |
The note ends with the sentence that stops a futile re-run:
Cleanup repairs occlusions; it does not invent frames that were never recorded, so re-cleaning will not fill these.
It also says where the holes are, in a line such as At 00:00:12, 00:01:44 and 2 more., so that a hole you can see on screen can be matched to the note.
Open frames inside a repair
Separately from the three kinds above, a repair itself can leave frames open. This is now rare: where the usual repair of a gap gives up, Post ships a best-attempt fill and says so on the card. A gap stays open when every fill Post could make would move the marker faster than a real body part can move.
- A region card reading Nothing filled shipped no value for that span, and it says why in plain words. You may read The rebuilt frames moved the marker faster than a real body part can move, so Kapnetix did not use them., or The rebuilt frames did not line up with the captured frames on either side, so Kapnetix did not use them. That card carries no model attribution and no score, because nothing was fixed.
- A repair that shipped part of its span keeps its model and states both halves: the frames that landed, and the frames still open.
- The Cleanup Plan block counts what the whole delivered file still has open, once:
4 frames ship open. That number is measured from the delivered file. Where nothing measured it, the panel saysthe plan ran, but the open frames were not recordedrather than working one out. - The region list header counts the same thing per region:
168 regions repaired · 59 regions that need eyes · 1 reverted to raw.
A repair with no score
Every repair carries a score from the cleanup. Where a repair shipped whole but the cleanup sent no score, or a score of 0, the card does not show an empty bar. It says why instead, for example No score recorded. The cleanup did not send a score for this repair, so check it by eye before you export., or The cleanup scored this repair 0 because part of it was not used and the captured frames there were put back, so check it by eye before you export. The region counts as still needing work.
Revert
Every region card on Review has a Revert button. One click excludes that region from the Clean version, the card then reads Reverted to raw capture · excluded from Clean v1, and Undo puts it back. Clean v1 is the first cleaned version of a take; your raw capture sits under it, untouched. A reverted region counts as work still to do, because the file you would hand over has the original problem in it.
On export, a revert is an exact restore of your original capture over those frames. Nothing is recomputed, and the rest of the clean stands. The export dialog says which state it is in:
You reverted 2 regions, so this export puts your original capture back over those spans. Nothing is recomputed, and the rest of the clean stands.The reverted export is still being prepared, so a file downloaded right now would still contain every AI fill for the 2 reverted regions. This usually takes a few seconds.
After the export it says what the file you just got actually contains: 2 reverted regions restored to your original capture in this file.
A cut that Post measured and then put back has nothing to revert, so it has no button. The frames in the file are already yours.
What this page does not cover
- Which model filled a gap. The region card names it, and the full model id is in its tooltip.
- Why a step did not run at all. That is a plan question. See Step states.
- Spikes. A spike is cut and refilled rather than filled. See Jitter.
What to do next
Open a cleaned take on Review and read the note above the region list first. Then switch the region list from By frame to By confidence to put the shakiest repairs at the top. Revert what you disagree with before you export.
Ask us
Something looks wrong on a take? Write to support@kapnetix.ai and send the take id, or the take's name and the project it sits in.