Calibration generates bad focal length about blam

[offtools]

Hi,

It could be, that the problem appears when working with an aspect of 4:3 instead of 16:9.
We've tried the addon today with a projector (resolution 1024x768, aspect 4:3) and also got a wrong focal length.
Later at home I tried to reproduce the failure without a video projector but with the lens data of the projector (Focal length, lens shift). I rendered a scene with the optical data of the projector, did the calibration work based on the rendering and tried to match the scene with the calibrated camera with the rendering as background image. I found out, that the problem may only appear when working with an aspect ratio that differs from 16:9 (4:3 in my case), with an aspect of 16:9 it seems to work. Will try this soon as possible with a projector again.

I added a small test, where I tried to match the calibrated camera with camera background image of the used scene. It shows the problem and a workaround when working with 4:3 aspect videoprojectors.

[stuffmatic]

Sorry about the late reply. Could you share any of your test images?

[thegoodhen]

I can confirm this. Using 4:3 images generates invalid estimates of focal length. I have tried setting the aspect ratio of the camera to the aspect ratio of my image manually before running BLAM, but it didn't help. I am getting relatively consistent (wrong) results for the image with 4:3 aspect ratio, while getting correct results for 16:9 images. So far this does NOT seem to be due to numerical instabilities.

[thegoodhen]

Okay, I have looked into it and I have found the issue (I think).

In the original file there is a TODO saying:

    #TODO: make sure this works for all combinations of
    #image dimensions and camera sensor settings

Well, I have done it and it doesn't. It appears that cam.data.sensor_height is set incorrectly, which may be a Blender bug (I am not sure about it). The ratio of cam.data.sensor_width to cam.data.sensor_width appears to be 16:9 (1.77777) no matter the image aspect ratio.

I have found the following workaround (which appears to work):

Find the block starting with
if imageWidth >= imageHeight:

and replace the whole if-else block with:

    if imageWidth >= imageHeight:
        theAspectRatio=imageWidth/imageHeight
        theHeight=cam.data.sensor_width/theAspectRatio
        fMm=theHeight * f
        #fMm = cam.data.sensor_height * f
    else:
        fMm = cam.data.sensor_width * f

I would post a pull request, but it seems like the author is not active any more.

EDIT: I am not sure if one of the calibrated intrinsic parameters is a pixel aspect ratio. I do believe that in the original paper it is so. I cannot find it in the code anywhere. If this is the case, the code should be reviewed some more, as my workaround (probably?) assumes that the pixel aspect ratio is 1.

[Bigbadcat612]

EDIT: I am not sure if one of the calibrated intrinsic parameters is a pixel aspect ratio. I do believe that in the original paper it is so. I cannot find it in the code anywhere. If this is the case, the code should be reviewed some more, as my workaround (probably?) assumes that the pixel aspect ratio is 1.

Hi! I used your code but I get a focal length of 900 mm in a square image. So it didn't work at all. Maybe that caused because I used x and z-axis to determine the focal length? Sorry for bad English.

Here's the image:

[thegoodhen]

EDIT: I am not sure if one of the calibrated intrinsic parameters is a pixel aspect ratio. I do believe that in the original paper it is so. I cannot find it in the code anywhere. If this is the case, the code should be reviewed some more, as my workaround (probably?) assumes that the pixel aspect ratio is 1.

Hi! I used your code but I get a focal length of 900 mm in a square image. So it didn't work at all. Maybe that caused because I used x and z-axis to determine the focal length? Sorry for bad English.

Here's the image:

I'm unable to reproduce the described bug with my version of code (with changes I described above). With your image, it works fine for me, following the tutorial on:
https://www.youtube.com/watch?v=yeWzQYh7iH0

On Blender 2.49b 64 bit. I have however first (before following the tutorial) set the background image to your image and selected "fit" under background image options (stretch/fit/crop). Not sure if that makes any difference and I don't really have the time to double check. I will upload the solved .blend file for your convenience.

EDIT: Well, this is odd... I have now looked into it a bit more and the solution is unexpected to say the least. It fits, but the focal length is very low and the object itself is very long. But doesn't that imply the problem is underconstrained? It is confusing to me.

EDIT2: HMMMMMMMMMMMMMMM... wait.

I think that this is actually an unrelated problem. The plugin works by computing vanishing points and with almost-parallel lines, the position of the vanising point (i.e. their intersection) is highly numerically unstable. There is one alternative approach (that I've read a paper on), which uses assumptions about orthogonality as opposed to vanishing points and which yields a better numerical stability, but I don't think that's implemented in BLAM.

[thegoodhen]

The .blend file:
https://drive.google.com/file/d/1URezwZXfG-Bt2yUT7RSLXtmqVPw7-_JY/view?usp=sharing

My (edited) blam.py:

https://drive.google.com/file/d/1GPdXrPzSqt_CFsmowqsdNqwLugD5yRlp/view?usp=sharing

[Bigbadcat612]

The .blend file:
https://drive.google.com/file/d/1URezwZXfG-Bt2yUT7RSLXtmqVPw7-_JY/view?usp=sharing

My (edited) blam.py:

https://drive.google.com/file/d/1GPdXrPzSqt_CFsmowqsdNqwLugD5yRlp/view?usp=sharing

Well, thanks for the reply!

It seems for me best to look an alternative to BLAM. Maybe autodesk image modeller or smth.

[stuffmatic]

In case you've missed it, BLAM is no longer actively maintained. You may want to check out fSpy

[Bigbadcat612]

Wow, thanks! I missed fSpy for some reason. Gonna check it out

…

On Sat, Jan 5, 2019 at 10:59 AM Stuffmatic ***@***.***> wrote: In case you've missed it, BLAM is no longer actively maintained. You may want to check out fSpy <https://fspy.io> — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#4 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AcJv9_MqdTBjhouhgP-fP-82zq4BBG2Nks5vAGmLgaJpZM4AwhVB> .