Metashape 1.6 out of core method very slow with photos from certain camera

[LipcaCZ]

Hi,

I was using RC last few years, but as they totally fucked up their pricing model, I'm really thinking about going back to Metashape (plase don't make the same mistake like they did).
I've seen that since 1.6, metashape is now using similar Out-of-core method of reconstructing model like RC, which is great (and you should really make much more advertising about it, because this is the most important feature you have ever added). Great work!

So I wanted to try it, but first I was very disappointed by the speed of the reconstruction. I took dataset with around 70 photos taken with Sony A6000 and only depthmap computation took about 40 minutes (on high detail). That's basically no diffrence between out of core and older method without using ooc.
But then I tried different dataset with 115 photos from differrent cameras (Canon EOS 600D) and it was much faster - only about 5 minutes for creating depthmaps.
So finally I took third dataset of 70 photos taken with Sony A7R and it was very fast again - only about 10 minutes, which is pretty surprising, because A7R has bigger resolution than A6000 (32 Mpix versus 24 Mpix)

So for better clarity, here's recapitulation:

Sony A6000 (24Mpix) - 40 minutes
Sony A7R (32 Mpix) - 10 minutes
Canon EOS 600D (18 Mpix) - 5 minutes

Could you please take look at this dataset taken with A6000 and try to find, why is it so slow? It's not only case of this dataset, I tried even different one from same camera and it was really slow either, so it looks like everything from this camera will be slow.

Link for dataset download: https://mega.nz/#!asREzCCT!zN9Rry5aqbK0A-BdNf1xyNcPx1YLQzhfbateT8zEuUU

Thanks!

[Mak11]

LipaCZ

What is the subject/content of the photos shot with the 600D & A7R ? Is it similar to the A6000 (architecture/wall/home) ?
Usually the content being reconstructed can have an impact on the processing time.

Mak

[LipcaCZ]

That's a good point! With 600D it was person scan from camera rig. With A7R it was a T-34 tank . Now testing a wall, but not so rough like that one taken with A6000 and it looks it will take about 20 minutes (same pohots count). I can try to go to the same place like I scanned with A6000 and rescan it with A7R, it's nearby my home.

[Mak11]

LipcaCZ

Yes, scan of objects/persons etc usually take less processing time simply because they don't take up the whole image like for example when you scan a wall or floor to create a material (which in turns requires to use nearly all the pixels of the photos). Also, the bigger the reconstructed volume/bounding box, then more processing time.

Just in case, also note that Version 1.6.1 is currently acting "funny" on Ultra High: https://www.agisoft.com/forum/index.php?topic=11755.0

Cheers

Mak

[LipcaCZ]

OK, so that second wall taken with A7R just finished in about 25 minutes (only depthmaps creation, not whole model). This one wasn't so rough like the other taken with A6000, so it looks like there's nothing wrong with that photos from A6000 in that case.

Well, that's a little disappointment, because that means that Metashape is still several times slower than RC (which completes whole model in about 15 minutes, depthmaps only 5 minutes).

Thanks for your points Mak!

[Mak11]

Took me 35 Minutes to process your A6000 photos on High (i7 6700K/32GB/Radeon VII) without any tweaking.
Yes RC is still faster in some cases but once again it always depends on the subject being scanned (Metashape's 1.6 GPU texturing is lighting fast though & RC simply can't compete there).

But RC's pricing model is simply ridiculous now. So it depends on your needs. RC is also Windows & CUDA only while Metashape works on everything (Linux/Mac/PC with OpenCL & CUDA)

Mak

[LipcaCZ]

35 minutes only depthmaps or complete model?

[Mak11]

Complete model.

Depth Maps: 22mins
Mesh: 13mins

Mak

[JMR]

In my opinion is overlap what makes a diference and not the subject, camera brand or model (at least I cant see a single fact related to brand or model besides pixel count). Try to compare datasets where the structure is the same.
Please let us know what you find.

GEOBIT