 |
|
|
|
|
|
 |
| ||||||||
 |
 
|
|
|
|
|
|
|
|
|
| ||||||||
|
|
| 
| User list | You are not logged in.
Being able to place CPs manually is a recurrent forum subject, for several evident reasons ...
I assume you are admiting that the most important point in CPs control is to delete first any "stupid" link (for exemple between images that includes looking like but distinct objects) and to delete any "harmfull" link (for exemple a link between two corners of fisheye images.)
A problem to place CPs manually is the amount of CPs to be managed. Contrary to stitchers where CPs are not found automatically (that is, where 5 CPs per link is usually enough) Autopano default number of CPs per link is 50: such a large number would make manual control of individual CPs very tedious at best or impossible depending on the source images number. To place 5 well located CPs in a pair of images when the other links include 50 CPs would not work (this link would be neglected by the optimizer.)
A logical solution could be to use the Keep only the N best control points option where N is possibly as low as 5.
I just made a 5 cp/link versus 50 cp/link test: initial RMS of the 5 cp/link version was bad (RMS 9 versus RMS 4 for the 50 cp/link version) but when I removed the "stupid & harmfull" links in both versions the advantage was for the 5 cp/version: RMS 1.9 versus RMS 3.8 for the 50 cp/link version (other settings were identical.)
If, as a rule, it's possible to get good results with 5 or even 10 CPs per link then I would be tempted to ask for a rather basic tool to implore SIFT to do its best at placing a single control point in a very small region we would specify in both CP Editor windows.
--- French / français ---
Un sujet récurent pour plusieurs raisons évidentes...
Une fois admis qu'avec Autopano il vaut mieux commencer par suprimer les liens "idiots" s'il y en a (par exemple ceux qui sont basés sur un objet qui ressemble mais est différent) ou les liens "nuisibles" (par exemple ceux qui concernent les coins des images fisheye) la question du nombre de points devient un problème: contrairement aux programme concurents où les points de contrôle sont souvent tous placés à la main donc peu nombreux (avoir 5 points de contrôle est courant dans ce cas) le nombre de points est d'environ 50 (par défaut) avec AP et ça rend une intervention manuelle point par point très fastidieuse ou impossible selon le nombre de photos à assembler. Mettre 5 CPs même très bien placés dans un lien alors que les autres liens en comportent 50 ça ne marcherait pas (ce lien serait presque ignoré par l'optimiseur.)
Une solution logique serait donc d'utiliser l'option Keep only the N best control points qui permet de limiter ce nombre et de descendre jusqu'à 5.
Je vien d'essayer d'assembler un pano (fisheye, 12 photos, 360°) d'une par avec 5 CPs par lien et d'autre part avec 50 CPs par lien (autres paramètres identiques): au premier assemblage le résultat avec 50 CPs est bien meilleur que celui avec 5 CPs mais une fois les liens "idiots" supprimés (dont un ayant un RMS de 50!) c'est au contraire le pano avec 5 CPs par lien qui a le meilleur RMS.
S'il est possible d'avoir de bons assemblages avec 5 points de controle (ou même 10) par lien il devrait être possible de désigner dans chacune des fenêtre de l'éditeur une toute petite région ou un et un seul CP doit être choisi par SIFT sans que ça nécessite une interface aussi riche et complexe que celles des concurents (lesquelles sont le fruit de nombreuses années d'évolution.)
Offline
I agree with Georges in that manual CPs need to be treated differently than automatic CPs.
But I think the answer is to count the manual CPs more. Perhaps yet more configuration items:
item 1. A manual CP is worth xx automatic CPs (perhaps this should be the same as the number configured to be found per link?)
item 2. Remove bad CPs needs a checkbox to allow it to remove manually placed CPs (or a seperate button to do this).
item 2a. Need to be able to remove only automatic or only manual CPs in CP editor. (when you select a rectangle then remove CPs)
item 3. optimize: use only manual CPs when automatic and manual CPs are present in a link. This includes lens distortion.
item 4. use only manual CPs as key for RMS calculations when manual and automatic CPs are present. That is, Assume the correct image placement is based only on manual CPs. Calculate the RMS of automatic CPs based on the average of the manual CPs. -- if there is 1 manual CP, it has RMS 0 and the RMS of the automatic CPs are calculated based on this. If there are two or more manual CPs, do the RMS calculation on these CPs only, then find the "ideal" picutre placement and calculate the RMS of the automatic CPs based on that.
-----
Don't know if this idea is good or bad:
Set 5 manual CPs at once. APP would show 5 CP candidates, four at the edges of image overlap and one in the center (like the 5 on either die of a pair of dice) The user would then refine these to be at the corners of the overlap, and one near the center. Note that if two or more auto CPs are found, APP knows where to put the 5 CPs automatically, if no auto CPs are found, the user has to select one or two matching features.
Last edited by hankkarl (2009-07-02 19:23:27)
Offline
I hope the next version of AP consider that. I think and it's obviously the manual control point is more important that automatically CP; then, the weight of these control points has to be more significant.
I asked about this in another post, but not about the value or importance of these.
Bear in mind it's critical to the successful development of a landscape.
Offline
ahornero wrote:
Bear in mind it's critical to the successful development of a landscape.
Why?
Offline
ahornero wrote:
Bear in mind it's critical to the successful development of a landscape.
The truth is that having no manual CPs compeled Kolor to implement some very interesting features like:
- different zoom settings in the same pano (for landscapes you can shoot the sky using a wide angle lens, shoot details using a long lens and stitch them
- manual moving/rotating mode of source images
- templates
- certainly many more but less visible ones...
My above suggestion corresponds (more or less) to an advanced feature of Max Lyons' PTassembler where the user click on a (well selected) pixel in one image and the program attempts (and often succeds) in selecting the coresponding location in the other image of the displayed images pair. Whether SIFT could or could not be used for that, I don't know!
BTW, I'm still convinced that managing a large number of CPs (1000 CPs is something very usual, much more is often encountered) is an impossible task. AP includes several tools to delete links, several tools to delete misplaced clouds of CPs (moving parts or identical features), several tools to add clouds of CP where the user find oportunities which could have escaped to SIFT.
To my surprise the optimizer seems to work with a very low number of CPs per link as well, though this still needs to be confirmed...
Last edited by GURL (2009-07-06 12:57:59)
Offline
GURL wrote:
My above suggestion corresponds (more or less) to an advanced feature of Max Lyons' PTassembler where the user click on a (well selected) pixel in one image and the program attempts (and often succeds) in selecting the coresponding location in the other image of the displayed images pair. Whether SIFT could or could not be used for that, I don't know!
Hugin uses this method, I found it very useful the few times I used it on panos that I could not stitch with APG because APG as of yet has no manual CP placement. You're probably wondering what an example case might be, well it was a case of stitching an IR image with a normal image. Both were sharp, but because of the obvious differences between IR and visible light, APP-1.4 failed to detect any CPs.
Hugin can also use SIFT. Would be nice if the manual placement of CPs in AP also included an automatic attempt to place the CP in the correct image (I place one CP on image A, AP tries to place it in the correct spot on image B), and we could then manually move it if its not where we want it. It saves time.
Offline
I've used this feature in Hugin a while back, and it seemed that the first CP was a wild guess, but the second and subsequent CPs used an offset from the first CP. That is, when you placed the second CP on image A, Hugin guessed about where the second CP should go on image B, and then refined it automatically. Sometimes it worked, sometimes it didn't.
Offline
ahornero wrote:
Bear in mind it's critical to the successful development of a landscape.
I've done plenty of landscapes with APP and never wanted manual CPs.
Interior shots, on the other hand, do cry out for manual CP placement.
Offline
Interesting reading. Really.
- "control point between IR and visible light" : If I could make SIFT compatible with that, it would make me really rich. That's the holy graal of medical imagery. No robust solution exists yet.
- I agree about the fact that manual CP should have more power than automatic one. Thus I don't know how to code that ... We could do that on a link easily but not on individual CP. Imagine that for an image pair, you can both kind of CP.
Offline
How about counting the manual CP as 10 (or some user-configurable number) of automatic CPs?
Now (I think) a link is positioned at the average (error? shift?) of all CPs
With manual CPs, the link could be the average of all CPs, but each manual CP would be added in 10 times.
Last edited by hankkarl (2009-08-25 22:12:51)
Offline
The mamual CP may solve the problem with noisy images where automatic detection puts CP between any images just on noise. I note also that the CP most of the time are not on sharp well conrasted edges, where arrors are annoying after assembly, but on more or less uniform surfaces. In such situation placing a manual CP would be salvatory.
Offline
- I agree about the fact that manual CP should have more power than automatic one. Thus I don't know how to code that ... We could do that on a link easily but not on individual CP. Imagine that for an image pair, you can both kind of CP.
Maybe its possible to add weighted cp`s by atomagically adding nearby cp`s when adding a manual cp ? The user may have the option to add 1, 4 or 12 cp`s.
User sets cp 1203/1484 and APP adds 1204/1285 etc...
Offline
hankkarl wrote:
I've done plenty of landscapes with APP and never wanted manual CPs.
Interior shots, on the other hand, do cry out for manual CP placement.
Gigapans from some height with water and a line of islands in the horizon often needs manual cp`s.
Offline
Pawel wrote:
The mamual CP may solve the problem with noisy images where automatic detection puts CP between any images just on noise. I note also that the CP most of the time are not on sharp well conrasted edges, where arrors are annoying after assembly, but on more or less uniform surfaces. In such situation placing a manual CP would be salvatory.
AFIK, that's what SIFT does--it avoids edges. IMO, APP should distribute some CPs evenly across the image, and should concentrate some CPs near edges, the higher the contrast, the more CPs should be placed near them.
Offline
JohnM wrote:
hankkarl wrote:
I've done plenty of landscapes with APP and never wanted manual CPs.
Interior shots, on the other hand, do cry out for manual CP placement.Gigapans from some height with water and a line of islands in the horizon often needs manual cp`s.
Ahhh....but that's a seascape 
Offline
