NEUTRALHAZER
PAPYWIZARD
KOLOR EYES
AUTOPANO VIDEO
PANOTOUR
AUTOPANOControl Point Editor: how to use it efficiently?
12 posts
Page 1 of 1
-
GURL - Member
- Topic author
- Posts: 2943
- Likes: 0 post
- Liked in: 0 post
- Joined: Tue Dec 06, 2005 1:57 pm
- Location: Grenoble
- Info
Control Point Editor: how to use it efficiently?
(Not sure about the right forum: there are issues related to documentation but some changes are suggested, too...)
In a few words: I find the question "how to use this tool efficiently ?" not being very easy to answer (steep learning curve!)
As I understand it, a user can make changes:
- at image pairs level (link after link) using Add Control Points tool and Delete Control Points tool. Removing all CPs between two images is equivalent to delete the corresponding link (deleting the link is faster.)
- on the whole panorama in a single step using Add control points using geometrical analysis = Automatically Add Control Points tool and using Clean bad control points = Automatically Delete Control Points (the names being used in the documentation draft don't correspond to the names of the actual tools in RC3.)
Panorama optimization is always a global one, I found no way no way to optimize a single image placement while leaving the other images unchanged (?)
The ways I was able to use CP editor with good results are:
1) to completely remove a link that should not exist (either 'stupid' CPs were created or CPs were placed in highly distorted regions like corners, etc)
2) to create a new link where one is needed
3) to remove CP at places like pano-head or moving clouds/people/cars where none are wanted
4) to add more CPs between two linked images at places where no CPs exist yet
5) to remove all CPs (pertaining to any link) whose 'RMS' is larger than a specified value.
As a PanoTools user, I'm rather oriented toward adding/moving/removing a small number of CPs and somewhat confused when having to manage such a large number of them!
My attempts to use Add control points using geometrical analysis resulted in many low quality CPs being added in the corners of very wide-angle images and on movings clouds: I suppose its usefulness depends on the images being used and of more appropriate options settings ?
Notes:
- About Customizable auto fine tune = Automatically add and delete links/control points: though I made several attempts using various option settings, I was not able to make any change using this button...
- It looks like an Auto Level action could be applied automatically by APP after each click on Optimize the panorama because pitch and roll are always re-optimized (in other words, any previous adjustment using vertical lines or any other tool is lost.)
- Sometimes a manual 90° or 180° rotation is needed after Optimize the panorama: this is certainly for the same reason as above.
- Selecting a small region on two corresponding source images results in a high concentration of CPs in this region: what if the CP number is much higher between two images than it is in the other images pairs?
I hope this helps!
In a few words: I find the question "how to use this tool efficiently ?" not being very easy to answer (steep learning curve!)
As I understand it, a user can make changes:
- at image pairs level (link after link) using Add Control Points tool and Delete Control Points tool. Removing all CPs between two images is equivalent to delete the corresponding link (deleting the link is faster.)
- on the whole panorama in a single step using Add control points using geometrical analysis = Automatically Add Control Points tool and using Clean bad control points = Automatically Delete Control Points (the names being used in the documentation draft don't correspond to the names of the actual tools in RC3.)
Panorama optimization is always a global one, I found no way no way to optimize a single image placement while leaving the other images unchanged (?)
The ways I was able to use CP editor with good results are:
1) to completely remove a link that should not exist (either 'stupid' CPs were created or CPs were placed in highly distorted regions like corners, etc)
2) to create a new link where one is needed
3) to remove CP at places like pano-head or moving clouds/people/cars where none are wanted
4) to add more CPs between two linked images at places where no CPs exist yet
5) to remove all CPs (pertaining to any link) whose 'RMS' is larger than a specified value.
As a PanoTools user, I'm rather oriented toward adding/moving/removing a small number of CPs and somewhat confused when having to manage such a large number of them!
My attempts to use Add control points using geometrical analysis resulted in many low quality CPs being added in the corners of very wide-angle images and on movings clouds: I suppose its usefulness depends on the images being used and of more appropriate options settings ?
Notes:
- About Customizable auto fine tune = Automatically add and delete links/control points: though I made several attempts using various option settings, I was not able to make any change using this button...
- It looks like an Auto Level action could be applied automatically by APP after each click on Optimize the panorama because pitch and roll are always re-optimized (in other words, any previous adjustment using vertical lines or any other tool is lost.)
- Sometimes a manual 90° or 180° rotation is needed after Optimize the panorama: this is certainly for the same reason as above.
- Selecting a small region on two corresponding source images results in a high concentration of CPs in this region: what if the CP number is much higher between two images than it is in the other images pairs?
I hope this helps!
Georges
-
ssprengel - Member
- Posts: 34
- Likes: 0 post
- Liked in: 0 post
- Joined: Fri Sep 01, 2006 3:20 pm
- Location: Nebraska, USA
- Info
On first glance, your message would appear to belong in the Support forum because it asks for clarification about the software's operation; however, in this case, you are commenting about features only available in the Beta versions so it probably belongs in the Development forum, because the Support forum is only for the released versions. The Future Requests forum is for requesting new features or enhanced operation, so I'm pretty sure the message does not belong here, but I am not any sort of moderator or list owner, and will leave it alone.
Aside from your observation that the Link Editor is difficult to learn how to use effectively (where I would characterize it more as some panoramas are difficult to fix and you really only have to know that you should use Optimize after adjusting things), many of your observations agree with mine, though I have never used PanoTools so can't speak to the differences between the two, nor am I some expert in APP, just a reasonably satisfied user waiting for the current release candidate to have all the bugs removed, but there are some things to comment about:
I usually use the Delete Control Points icon in the Link Editor to remove SOME of the control points that appear to be in the wrong places just before using the Add Control Points icon to have it compute new ones in the little areas I've selected. Or sometimes I delete all existing control points and then add all my own, but if I want to remove a wrongly linked image, I would just delete an entire link rather than deleting all the control points, although either way would work, as you say.
If you want to add control points (using geometric analysis?) across just two specified images, you can bring up the image pair in the link editor and use click-drag to highlight each image entirely (as opposed to two small areas that seem to correspond), and use the Add Control Points icon up top (the one with a red plus amidst a cloud of pluses) to have APP figure out what corresponds between the two. I am not recommending you do this for the same reasons as you may not want to do overall geometric analysis because you may get many bad control points, but it can be useful, sometimes.
Optimization adjusts the image positions of the entire panorama to minimize the global Root Mean Square (RMS) of the panorama, which could change even with a single change in a control point, so yes it is a global thing. Individual images are never locked in place compared to others, so moving one image in relation to the rest of the pano is not really how you want to think about what's happening.
Depending on what's been done with links and control points, optimizing can completely change the geometry of the panorama so it does forget the existing manual adjustments to yaw pitch and roll as they may no longer apply, and for the same reason, you must usually redo the auto-level as well as respecifing what are at least three verticals to fine-tune unwarping the horizon. I say this because I typically get a curved horizon after optimization. After optimization, the center of the panorama also is likely a different spot. So, in general, if you find yourself messing with rotation and straightening alot, then do all your control-point editing and optimization, first, and then straighten and rotate things...because optimization forgets anything that depended on the previous geometry. About the only thing it doesn't forget, which I wish it would, half the time, is the crop box, since I tend to see lots of transparency after an major optimization, and have to remove and reapply the cropping. Obviously cropping before tweaking control points is not the most useful thing, either, usually.
I typically would use Clean Bad Control Points after adding my own control points or after using geometric analysis to add control points, just to cut down on the computations at later stages. However I generally don't clean control points at every link adjustment because what are bad control points at one point may be good ones after I've reoptimized. It really just depends on how messed up things are and if you have to do alot of tweaking to get things right.
It is probably not a good idea to use Geometric Analysis to add control points if the images aren't very well matched, because, as you say, you will likely get bad ones, but depending on how lazy you are or if you just want to see if the computer can do better than you, it can sometimes be fun and sometimes useful to set the control point adding settings to not replace old with new but just add to existing control points, and then do several cycles of: Add with Geometric Analysis, Optimize, Remove Bad Control Points, so see if things improve or get worse. You can always roll back a few steps in the history if things are steadily getting worse. You probably don't want to do this with a large panorama as it will take quite a bit of time to recompute new control points based on geometry when there is quite a bit of image overlap. I also will have to take your word for it that Remove Bad Control Points is panorama wide, as I've barely used it, and recently on just a two-image panorama.
I am also mystified by the Auto-Fine-Tune icon, since it doesn't seem to do anything for me, either. I have not used APP long enough to know what it used to do, if it did do something in the past, or if I haven't quite set things up correctly to have it do anything.
I also have wondered what increasing the number of control points between two images does. I would guess it either does nothing, or more likely gives that image pair more weight in the overall optimization...I haven't thought about it hard enough, nor read through all the messages and wiki to convince myself one way or the other.
Aside from your observation that the Link Editor is difficult to learn how to use effectively (where I would characterize it more as some panoramas are difficult to fix and you really only have to know that you should use Optimize after adjusting things), many of your observations agree with mine, though I have never used PanoTools so can't speak to the differences between the two, nor am I some expert in APP, just a reasonably satisfied user waiting for the current release candidate to have all the bugs removed, but there are some things to comment about:
I usually use the Delete Control Points icon in the Link Editor to remove SOME of the control points that appear to be in the wrong places just before using the Add Control Points icon to have it compute new ones in the little areas I've selected. Or sometimes I delete all existing control points and then add all my own, but if I want to remove a wrongly linked image, I would just delete an entire link rather than deleting all the control points, although either way would work, as you say.
If you want to add control points (using geometric analysis?) across just two specified images, you can bring up the image pair in the link editor and use click-drag to highlight each image entirely (as opposed to two small areas that seem to correspond), and use the Add Control Points icon up top (the one with a red plus amidst a cloud of pluses) to have APP figure out what corresponds between the two. I am not recommending you do this for the same reasons as you may not want to do overall geometric analysis because you may get many bad control points, but it can be useful, sometimes.
Optimization adjusts the image positions of the entire panorama to minimize the global Root Mean Square (RMS) of the panorama, which could change even with a single change in a control point, so yes it is a global thing. Individual images are never locked in place compared to others, so moving one image in relation to the rest of the pano is not really how you want to think about what's happening.
Depending on what's been done with links and control points, optimizing can completely change the geometry of the panorama so it does forget the existing manual adjustments to yaw pitch and roll as they may no longer apply, and for the same reason, you must usually redo the auto-level as well as respecifing what are at least three verticals to fine-tune unwarping the horizon. I say this because I typically get a curved horizon after optimization. After optimization, the center of the panorama also is likely a different spot. So, in general, if you find yourself messing with rotation and straightening alot, then do all your control-point editing and optimization, first, and then straighten and rotate things...because optimization forgets anything that depended on the previous geometry. About the only thing it doesn't forget, which I wish it would, half the time, is the crop box, since I tend to see lots of transparency after an major optimization, and have to remove and reapply the cropping. Obviously cropping before tweaking control points is not the most useful thing, either, usually.
I typically would use Clean Bad Control Points after adding my own control points or after using geometric analysis to add control points, just to cut down on the computations at later stages. However I generally don't clean control points at every link adjustment because what are bad control points at one point may be good ones after I've reoptimized. It really just depends on how messed up things are and if you have to do alot of tweaking to get things right.
It is probably not a good idea to use Geometric Analysis to add control points if the images aren't very well matched, because, as you say, you will likely get bad ones, but depending on how lazy you are or if you just want to see if the computer can do better than you, it can sometimes be fun and sometimes useful to set the control point adding settings to not replace old with new but just add to existing control points, and then do several cycles of: Add with Geometric Analysis, Optimize, Remove Bad Control Points, so see if things improve or get worse. You can always roll back a few steps in the history if things are steadily getting worse. You probably don't want to do this with a large panorama as it will take quite a bit of time to recompute new control points based on geometry when there is quite a bit of image overlap. I also will have to take your word for it that Remove Bad Control Points is panorama wide, as I've barely used it, and recently on just a two-image panorama.
I am also mystified by the Auto-Fine-Tune icon, since it doesn't seem to do anything for me, either. I have not used APP long enough to know what it used to do, if it did do something in the past, or if I haven't quite set things up correctly to have it do anything.
I also have wondered what increasing the number of control points between two images does. I would guess it either does nothing, or more likely gives that image pair more weight in the overall optimization...I haven't thought about it hard enough, nor read through all the messages and wiki to convince myself one way or the other.
Last edited by ssprengel on Tue Nov 28, 2006 1:39 am, edited 1 time in total.
-
AlexandreJ - Kolor Team
- Posts: 5960
- Likes: 1 post
- Liked in: 0 post
- Joined: Mon Nov 14, 2005 4:56 pm
- Location: Francin, France
- Info
Hum, interesting reading. Some comments :
- Crop before control points edition is bad. I don't know if we have to automatically set the "auto fit panorama" when doing an optimization or not. After all the user asked for a crop : so if we remove it when using optimize, will he understand why without showing the result ?
We could also think of an algorithm that will reapply the crop area over the optimized panorama, but it could never be accurate enough and the user will have to adjust the crop, making this algorithm useless.
- Some tips I found when working with the control point editor :
Having more control point on a link that the average number of control point anywhere else is a kind a balancing scheme : it will give more power to this pair of picture and force more this pair to be stitched well even if it may give more error on other pairs.
This is good as you may choose which area are more important for you : along an edge, for a close object, etc. Just have a greater number of control point there to force the optimizer to take more care about this place over the rest of the panorama.
You can achieve this easily by using the control point cleaner, second method : just keep the 20 better control point for each pair. So you will be sure that the power of each pair is 20. Then add some control point for the pair you want to prefer. BTW : the number of control point added is in fact determined by the average of CP per link. In this case, you'll get 20 more control points.
The geometry analysis doesn't work well with 360x180 panorama. It often try to determine too much overlapping zone on pole where it should not. This method really works on partial hand held panorama, where everything is placed randomly. The result panorama is more strongly linked and gives a better overall result. Now, it can also be achieved by just changing the detection setting from normal to high. The links that were ignored in normal mode will be detected in high mode thus making this algorithm useless.
As ssprengel said : this algorithm needs to have a well estimation of picture location to work. Don't forget that.
- Crop before control points edition is bad. I don't know if we have to automatically set the "auto fit panorama" when doing an optimization or not. After all the user asked for a crop : so if we remove it when using optimize, will he understand why without showing the result ?
We could also think of an algorithm that will reapply the crop area over the optimized panorama, but it could never be accurate enough and the user will have to adjust the crop, making this algorithm useless.
- Some tips I found when working with the control point editor :
Having more control point on a link that the average number of control point anywhere else is a kind a balancing scheme : it will give more power to this pair of picture and force more this pair to be stitched well even if it may give more error on other pairs.
This is good as you may choose which area are more important for you : along an edge, for a close object, etc. Just have a greater number of control point there to force the optimizer to take more care about this place over the rest of the panorama.
You can achieve this easily by using the control point cleaner, second method : just keep the 20 better control point for each pair. So you will be sure that the power of each pair is 20. Then add some control point for the pair you want to prefer. BTW : the number of control point added is in fact determined by the average of CP per link. In this case, you'll get 20 more control points.
The geometry analysis doesn't work well with 360x180 panorama. It often try to determine too much overlapping zone on pole where it should not. This method really works on partial hand held panorama, where everything is placed randomly. The result panorama is more strongly linked and gives a better overall result. Now, it can also be achieved by just changing the detection setting from normal to high. The links that were ignored in normal mode will be detected in high mode thus making this algorithm useless.
As ssprengel said : this algorithm needs to have a well estimation of picture location to work. Don't forget that.
-
[bo] - Member
- Posts: 1226
- Likes: 0 post
- Liked in: 0 post
- Joined: Fri May 05, 2006 8:16 am
- Location: Bulgaria
- Info
That's close to that I do - when I have a specific feature in some of the images, I add more CPs there, so it has more "weight" in optimization. This way the panorama seems to "fit" around that feature, which is usually the focus of the whole shoot. Problems start to show the farther away from the "heaviest" part of the panorama, but those are easily fixable and not a problem really, as they don't occur in/on my main object.
In a general landscape panorama, I try to "even the weight" in all links, using the "Leave N best CPs" cleaner option.
In a general landscape panorama, I try to "even the weight" in all links, using the "Leave N best CPs" cleaner option.
Some of my panoramas, posted in the Autopano Pro flickr group.
-
Al - Member
- Posts: 98
- Likes: 0 post
- Liked in: 0 post
- Joined: Sat Jan 21, 2006 1:35 am
- Location: Ann Arbor, MI
- Info
Agreed with kankkarl. Alexander, could you explain a bit about the control point weighting algorithm? For example, Bo mentioned above that he adds more CPs to a specific feature so that feature ends up being 'weighted' more in the optimization process. This seems intuitively correct, but it would be helpful to understand a bit more about how the process actually works. In other words, if I select a small area in two overlapping images (say 5% of the area of each) I might get 10-100 extra CPs clumped in a small spot, what then is the effect of the now more heterogeneous distribution of CPs across the entire pano on the optimization process? Does that make sense?
-
hankkarl - Member
- Posts: 1284
- Likes: 0 post
- Liked in: 0 post
- Joined: Tue Feb 21, 2006 5:32 pm
- Location: Connecticut, USA
- Info
Al wrote:Agreed with kankkarl. Alexander, could you explain a bit about the control point weighting algorithm? For example, Bo mentioned above that he adds more CPs to a specific feature so that feature ends up being 'weighted' more in the optimization process. This seems intuitively correct, but it would be helpful to understand a bit more about how the process actually works. In other words, if I select a small area in two overlapping images (say 5% of the area of each) I might get 10-100 extra CPs clumped in a small spot, what then is the effect of the now more heterogeneous distribution of CPs across the entire pano on the optimization process? Does that make sense?
I think, from what Alexandre said, that all CPs are equally weighted, and the optimization tries to determine a best fit by lowest overall error. So the number of CPs on or near a feature give that feature more weight.
Alexandre gave an example using 20 CPs per link, if took one link and added 80 CPs (total 100 CPs on that link) I think it would count five times as heavily when the optimizer runs. So the effect is that that one link will be better, but all the other links may be a bit worse.
But what about lens distortion? how does that play into the equation? Would a large number of CPs in one area affect it? I can see a couple of pictures overlapping in the corner having a large number of CPs affecting the optimization, but if the lens distorted near the corners, this could be a problem. Or does optimization weight points near the center of the lens more heavily?
-
AlexandreJ - Kolor Team
- Posts: 5960
- Likes: 1 post
- Liked in: 0 post
- Joined: Mon Nov 14, 2005 4:56 pm
- Location: Francin, France
- Info
The location and lens distortion will be affected in the same way by this weight.
Now, for lens distortion, a CP far from center will have more influence than a CP near of it, because distortion is a radial function.
BTW : what's why you have to disable lens distortion on a 2 pictures stitch that have the same center : it just gives infinite solution and the optimizer doesn't converge.
Now, for lens distortion, a CP far from center will have more influence than a CP near of it, because distortion is a radial function.
BTW : what's why you have to disable lens distortion on a 2 pictures stitch that have the same center : it just gives infinite solution and the optimizer doesn't converge.
-
ssprengel - Member
- Posts: 34
- Likes: 0 post
- Liked in: 0 post
- Joined: Fri Sep 01, 2006 3:20 pm
- Location: Nebraska, USA
- Info
Is there a way to disable lens-distortion after a panorama has been detected, as when I'm in the link-editor and trying to fix things? Sometimes panos look ok, initially, but then after some link-edits, they go all loopy at the edges, and the solution is to turn off lens-distortion and redetect. Could there be a less wasteful way of turning off lens-distortion than to start all over with detection?
-
Al - Member
- Posts: 98
- Likes: 0 post
- Liked in: 0 post
- Joined: Sat Jan 21, 2006 1:35 am
- Location: Ann Arbor, MI
- Info
Would it be possible for the CP editor to place new CPs in the selected area in number proportional to the CP density in the rest of the pano so as to not over-weight the spots with CPs placed through the editor vs those generated in the original pano?
I am using pro to stitch together shots from a digital microscope. Sometimes I change the focal length to better see detail in the background or foreground. The only way I can seem to gI'm surprised at the number of obvious matcheset to fit into one image everything is to manually add control points to most images and to further expand links to more than one other photo. Even then It seems like lots of obvious matches are being missed. Can someone suggest the best settings for this type of operation? Marty
-
klausesser - Member
- Posts: 8143
- Likes: 0 post
- Liked in: 5 posts
- Joined: Mon May 22, 2006 12:18 am
- Location: Duesseldorf, Germany
- Info
MartyB wrote:Can someone suggest the best settings for this type of operation? Marty
My first idea: don´t change focus length.
best, Klaus
Simplicity is the keynote of all true elegance. Coco Chanel
12 posts
Page 1 of 1
Who is online
Users browsing this forum: No registered users and 2 guests