Low GPU use on 1.6.0

[RoryG]

Hi,

I'm working on a 19,000 image project on 1.6.0 release version, I didn't want to update to 1.6.1 during a major project. I'm experiencing low GPU and high CPU use during depth map generation (see screenshots), in V1.5.x both GPUs were maxed out on CUDA and there was some breathing space on the CPU. Use CPU is not checked in the GPU tab. System is i9 9900X, 128GB RAM, 2 x RTX2080.

[Alexey Pasumansky]

Hello RoryG,

Are you processing the project where "fit additional corrections" option has been used for optimization?

[RoryG]

Hi Alexey, yes, at the last stages of gradual selection, I checked all the optimisation parameters after tightening the tie point accuracy to 0.3. I then did gradual selection on reprojection error down to 0.2 and optimised again checking all the boxes. I'm now building the dense clouds in duplicated chunks with around 1,000 images in each bounding box. Maybe I shouldn't be using the fit additional corrections option? The camera is an eBee SODA.

[RoryG]

Hi, I just read Dense cloud generation time longer, increased noise post and see it's the same issue. I haven't looked at any of the dense clouds yet so don't know if they are noisy or not. This section of the project has 5669 cameras and as they were all aligned and optimised together I've no option but to continue processing the chunks. The question is, when I do the other two sections without using the additional corrections option, will there be a discernible difference in the dense clouds when I merge all the projects together?

[Alexey Pasumansky]

Hello RoryG,

It is expected behavior (see post here: https://www.agisoft.com/forum/index.php?topic=11722.msg52609#msg52609) that the depth maps generation takes longer time for the cameras with additional corrections enabled in versions 1.6.0 and 1.6.1.

If the additional corrections are the only difference between the camera calibration models in the chunks, then I think, the difference shouldn't be observed. However, other calibration parameters may be also different and in such case the chance of having some misalignment between the processed clouds is higher.

[RoryG]

Interesting observation, the last chunk I was processing contained images from the SODA camera plus a section that was flown using the P4RTK. It took around 14 hours to create the dense cloud from around 1200 images. The chunk I'm processing now has only images from the SODA camera and is processing much faster, although the GPU use isn't maxed out but it is a fair bit higher.

[RoryG]

The project that used additional corrections took 36 hours to process 5.6K images with high settings, the second project without additional corrections took just 19 hours to process 6.2K images with high settings!