Agisoft Metashape

Agisoft Metashape => Bug Reports => Topic started by: juanmarquez1988 on February 28, 2014, 03:47:12 AM

Title: Same photo set, huge difference in results
Post by: juanmarquez1988 on February 28, 2014, 03:47:12 AM

Hello every one,  :)

Has any one had the same issue? ???
I've used the exact same photoset and processed it *3 times*, in the same conditions with the same optionsfor the initial alignment.
- Accuracy: High
- Pair selection: disabled
- Point limit: 40,000
- Constraint by mask: no

Oh, and the 28 pictures are 960x1280 and are well overlapped.  No blur. :)

I've only proceeded with the first step: alignment, and the camera locations is way way..wayyy different!  :o
I've struggled to come up with an explanation and I know it sounds weird, but I can't figure out what happened.  :-\

I hope somebody can help,
:-)

J.

Title: Re: Same photo set, huge difference in results
Post by: Alexey Pasumansky on February 28, 2014, 10:06:08 AM

Hello juanmarquez1988,

Could you post any original image? I assume that the difference in results may be caused by low number of feature points that can be found on the objects' surface.

Title: Re: Same photo set, huge difference in results
Post by: Wishgranter on February 28, 2014, 11:42:38 AM

what camera you used ? some low res one as see from  img resolution, Can post some screenshot ?

Title: Re: Same photo set, huge difference in results
Post by: juanmarquez1988 on February 28, 2014, 01:54:07 PM

Hi Alexey, Wishgranter,

Thanks for your response.
Here are the screenshots. I've used my Samsung Galaxy S3 camera,do you think it's hopeless wit it?

Title: Re: Same photo set, huge difference in results
Post by: Alexey Pasumansky on February 28, 2014, 02:04:41 PM

Hello juanmarquez1988,

The resolution is really low, but in principle you can get acceptable results with it.

In case you are taking images of the living subject, you should keep in mind that even little movements and geometry shape changes could lead to camera alignment problems.
Another important point is providing good lighting conditions, since blurred and noisy images are not any good for processing.
Also check if cameras are not marked by NC label in the Workspace pane, otherwise it is better to input initial values for camera calibration in the Camera Calibration window (focal length and sensor pixel size - both in mm).

Title: Re: Same photo set, huge difference in results
Post by: juanmarquez1988 on February 28, 2014, 02:19:46 PM

Thanks Alexey.  :)

Yeah I also figured out that resolution, light, noise, blur, and camera calibration might be the culprit for a poorer reconstruction.
I'm using a phone (S3), so the pixel size wasn't automatically recognized. However, the focal length is correctly detected.

However, the thing I don't understand is that: when I launch the pictures camera alignment for a single set of photos and with the same settings, I obtain radically different results.

Title: Re: Same photo set, huge difference in results
Post by: Alexey Pasumansky on February 28, 2014, 02:47:40 PM

The results of Align Photos stage may be slightly different due to the algorithm used.

Big differences means that PhotoScan fails to find the stable alignment results and due to low number of matching points may use invalid matches to align cameras.

Title: Re: Same photo set, huge difference in results
Post by: juanmarquez1988 on February 28, 2014, 02:52:35 PM

Okay, I'll try to get better pictures to put in PhotoScan.

Cheers
 :)

Title: Re: Same photo set, huge difference in results
Post by: Octavian on March 04, 2014, 01:37:30 PM

Trying out the demo and i've run into the same problem.
Though in my case the resolution and camera are definitely superior.
Using a canon 20d at 8.2 mpix (3500x2330 px) and getting very inconsistent results between the same alignment sessions. Sometimes it works perfectly, other times it fails completely.
The focal lenght and pixel size are also automatically detected correctly.
Sometimes i run the alignment with pair preselection set to disabled, and it keeps failing, then i suddenly change that to generic and i get a perfect result. Then i run it again with it set to disabled, and now it works, even though, before it was failing.
Sometimes i close the application, and when i start it again, it gets a perfect result from the first try.
Other times, i try it 10 times in a row and it keeps failing, with the exact same settings and images.

Wish i could understand a little better what happens under the hood, and why at times i get a perfect result all of a sudden. Specular reflections might be a cause of failure of course, with light moving across the surface of the barell, but in that case, shouldn't it fail every time?

Here's some pics to better depict the problem. As you can see in the couple of example images i've attached (40 in total), they are pretty well lit and i've also used some square markers hoping that it would help with the alignment.

Any suggestions would be highly appreciated,

Thanks,
Octav

Title: Re: Same photo set, huge difference in results
Post by: Wishgranter on March 04, 2014, 01:46:06 PM

try to have model on say 80+ % of the image, the problem with align on this dataset is because it get points from surrounding area and that can cause problems. when trying to scann this sort of objects, use newspaper under the object, taht AGI can detect mor epoints in area of interest and not in wide area around the object of scaning.....  jsut from seeing you area of interes - reconstruction area even with the platfrom occupy just approx 20-30 % of the image.... lets try it this way and let us know......

Title: Re: Same photo set, huge difference in results
Post by: Octavian on March 04, 2014, 04:20:33 PM

Thanks for the tips.
Will try that and let you know how it goes.
Removing the background and running the alignment seems to create pretty stable results,
so my guess is you were right.