Hugin

align_image_stack speed

Asked by Lachlan Taylor

Hi,

I am using align_image_stack to align some photos that I am later stitching together to make a short video. The tool seems to do what i require perfectly however the processing can take some time. I have decided to move the processing to the cloud and am now using a high powered GPU server with 2xNvidia graphics cards. However it still is taking longer than I would prefer (there is quite a bit of processing I need doing).

Can someone please tell me whether I need to set any particular setting on align_image_stack to use the GPUs for processing? I noticed fromm this version - http://fossies.org/linux/privat/hugin-2014.0.0_rc1.tar.gz/hugin-2014.0.0/doc/align_image_stack.pod - that there are two options that might help, --gpu and --threads= however these options don't seem to be available on my version of the tool. Is this because the newer version does this by default or should I compile the linked version instead and use that? Is there anything else I need to make sure is done to take advantage of the GPUs?

Thanks very much in advance.

Lachlan.

Question information

Language:
English Edit question
Status:
Answered
For:
Hugin Edit question
Assignee:
No assignee Edit question
Last query:
Last reply:
Revision history for this message
tmodes (tmodes) said : 		#1

The switches are only available in Hugin 2014.0 (currently RC2) and later.

Some comments: Align_image_stack is using by default the CPU for remapping and is using as much threads as cores (versions before 2014.0 were only using one thread). The number of parallel threads can be fine-tuned by the --threads switch.
The GPU can be used for remapping by specifying --gpu. In this case only one thread is used (the --threads switch has no effect in this case) and all remapping is done on the GPU. (But this does not work with all graphics systems: graphic card+driver. )

If speed is the main concern have a look on the default branch in the repository. Align_image_stack has got some more optimisations (but not released yet and not tested so much). For maximum speed you will need additional the libfftw3-header and -libraries.

Revision history for this message
Lachlan Taylor (lockytaylor) said : 		#2

Thanks for that tmodes... I'll check it out. How do I know which version I am using?

Revision history for this message
tmodes (tmodes) said : 		#3

In Hugin in the about dialog.
Or call align_image_stack --help

Revision history for this message
Lachlan Taylor (lockytaylor) said : 		#4

Thanks very much for that. I appreciate your help.

I managed to compile the 2014 version however it bombs out with a 'failed to open display' error when using --gpu option. I am using Nvidia GRID K520 graphic cards and Amazon Web Services' linux Nvidia driver. Apparently it will work with GRID and/or tesla M-class drivers. Do you know whether align_image_stack should work with either of these drivers? More info: http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/using_cluster_computing.html

Also, I am currently pushing the processing off to the cloud with AWS however I would interested in using one of these instead http://devblogs.nvidia.com/parallelforall/jetson-tk1-mobile-embedded-supercomputer-cuda-everywhere/. Do you know whether align_image_stack would work well with the specs provided in the link? Is align_image_stack CUDA compatible?

I am a bit of a newb a this stuff so please accept my apologies for any dumb statements... :)

Revision history for this message
Lachlan Taylor (lockytaylor) said : 		#5

bump

Revision history for this message
tmodes (tmodes) said : 		#6

> Do you know whether align_image_stack should work with either of these drivers?

There are no reports which graphic systems (card+driver) work and which one does not work. Currently there are only some older reports for desktop systems: http://wiki.panotools.org/Nona_GPU_stitching_reports

> Is align_image_stack CUDA compatible?
Align_image_stack is not using CUDA code. All the interaction with the graphic card is done via the OpenGL API.

Can you help with this problem?

Provide an answer of your own, or ask Lachlan Taylor for more information if necessary.

To post a message you must log in.