Agisoft PhotoScan 1.3.0 pre-release

[andrey.d]

Hello andrey.d,

For me it looks like the camera orientation angles loaded to the Reference pane is incorrect - so the model orientation in space in also incorrect. I've input very large number as "camera accuracy (deg)" to the Reference pane settings and then Updated the Reference pane - the orientation of the model after that looked valid.

I have image_name, longitude, latitude, altitude, yaw, pitch, roll
Could you give right format of csv file in my case?

[Alexey Pasumansky]

Hello andrey.d,

I can suggest to skip loading yaw, pitch and roll, or to use the following lines to lower the accuracy of the angles after you load the CSV:

Code: [Select]

chunk.camera_rotation_accuracy = (1E5, 1E5, 1E5)
chunk.updateTransform()


[Mak11]

Just grabed 1.3.0 do test it out and things are definitely broken. My photos are simply not aligning correctly at all (this is with OpenCL enabled. (in some cases the point cloud is sliced in half an translated to a different axis...wth) I also noticed that a big chuck of the photos are ignored in the process. Just retest in 1.23 and everything works fine. will do some testing without OpenCL now.
I'm attaching the Log and some screens of my second run where the "slicing wasn't as severe as the first time where literally the whole point cloud was perfectly slice in half vertically and shifted on the X axis)

[Mak11]

And here's the log attached.

[Mak11]

Just ran the alignment without OpenCL and things aren't better. A handful of cameras are not correctly aligned at all and a big chuck of the model is in the air..but this time only 2 of the 509 photos were dismissed (way less than with OpenCL). The second screenshot is how it looks in 1.2.6 (all cameras should be at ground level etc..)

[Mak11]

Ok, so I uninstalled 1.3.0 and re-installed 1.2.6. Opened the same project, hit align without changing anything and everything was re-aligned perfectly. Something definitely fishy in 1.3.0.

[Alexey Pasumansky]

Hello Mak11,

Can you provide the sample dataset were you are observing the problem to support@agisoft.com?

Also please specify if you are using exactly the same parameters for alignment.

[lunar]

i`ve got the same problem, with openCl poor align (1/3 of photos), turning off gfx card all photos aligned

[Alexey Pasumansky]

Hello lunar,

Could you specify the 1.3 build number and GPU model used?

[lunar]

1.3.3620  - old HD7800 pitcairn 1024MB - this maybe a problem - very old card

[Mak11]

Hello Mak11,

Can you provide the sample dataset were you are observing the problem to support@agisoft.com?

Also please specify if you are using exactly the same parameters for alignment.

Hello Alexey,

Yes as far I as could see I was using exactly the same parameters for alignement. I initially though that it was an OpenCL bug but the CPU only run was messed too (but a lot better than the OpenCL run) so alignment was incorrect in both modes while it's perfect every time I run it in 1.2.6.

The dataset is over 10GB of DNG photos :-( I don't think I could upload the whole thing in less than a week with my current internet connection.

[Alexey Pasumansky]

Hello Mak11,

Maybe the same issue can be reproduced on the subset of images for JPGs, for example?

[Mak11]

Alexey,

I'll try this out this week-end and let you know.

[Alexey Pasumansky]

Thank you Mak11,

Maybe you can also send the report files generated from 1.2.6 and 1.3.0 projects, just for case we can find something unusual there and try to reproduce the problem on similar dataset.

[Mak11]

How do I grab the report files? is it just the log (like the one I posted above?) or do I have to do anything else?