Max quality setting question

Hi there,
so the couple of days I was running a dataset but I’m still not satisfied with the results.
My question is like this: What are the balls-to-the-wall no-mercy absolute max settings you can set? I don’t care about processing time, nor RAM. I just want the max available quality.

cheers.

1 Like

With the warning that this is usually a Bad Idea™️:

--feature-quality ultra
--ignore-gsd
--matcher-neighbors 0
--min-num-features 64000
--pc-filter 1
--pc-quality ultra
--pc-sample 0

You can take it further by editing the config.py in ODM, but the likelihood of the above being more than anyone can reasonably process at home is high.

What are you dissatisfied with in your results? What does your data look like?

Thank you. As soon as my current workflow is finished, will try your settings (probably tomorrow)
I’m missing these pillars https://i.imgur.com/jkvSQvl.jpg even with ultra quality settings. The screenshot is from Pix4d…

1 Like

–feature-quality ultra
Yes
–ignore-gsd
DONT DO IT, you told me!
–matcher-neighbors 0
Bad idea! A max value can be calculated. Matcher distance can also be calculated. I know how! With value 0 you will calculate for ever in vain.
–min-num-features 64000
A bit to much, I say 16000 is good enuf.
–pc-filter 1
Depends on GSD!
–pc-quality ultra
Yes
–pc-sample
Depend on GSD!

1 Like

I know, and I told them it’s a bad idea and likely won’t process. But they asked 🤷 And yes of course, you should never use this option. It will not go well.

Easiest to set it in the options than to make someone do the math.

Depends upon the source material, but I have not really found a situation where more helped (though I have needed up to about 64k).

To an extent, yes, but relaxing it from the default of 2.5 can help retain some finer variations in surfaces.

Agreed, but again, they asked for the craziest options. Not filtering is one of them.

There’s only so much I can advise without knowing the data, and again, they wanted crazy.

They better ask for the intelligent settings.

I use ultra settings and filter for a GSD of 10mm to remove duplicates.

A wile ago I did an excel to calculate matcher-distance and matcher neighbours. Works really well and I know I will not loos time or quality.

But about min-num-futures, if the minimum isn’t found is a otherwise positive match thrown away?

1 Like

Fair point. Now we have this dialog out in the open perhaps we can advise them better :slight_smile:

I don’t believe so. My understanding is that it keeps relaxing various thresholds until desired number of features are tagged. It isn’t setting a threshold for number of valid matches, just valid features detected. Matching does not have exposed knobs for tuning outside of setting the matcher type currently.

I thought it was how many futures that needed to match between two images. I need to investigate!

1 Like

I would like to succeed tho, if it will fail for sure, then this is not an option.

1 Like

How much RAM do you have, how much swap/pagefile, what resolution images, and how many images?

Same goes for --pc-filter 0, it can easily get out of control. If you have an estimate of your GSD, you can set it to something like half of that to be a bit “safe” (remember, it is in meters, not centimeters like most other parameters).

So say you have an estimated GSD of 4.33cm/px (12MP at 120m/400ft AGL), you could pretty reasonably set it to something like --pc-filter 0.02 and not over-filter, nor under-filter and swamp your machine.

1 Like

For this current dataset 327x 17mpx images, about 2.4cm calculated GSD.
50 Gig RAM and 250Gig Swap. I can create 2TB more Swap if needed though.
Thanks!

1 Like

With Andreas’ sensible changes, you should be getting close to the cusp of what’s “best” without needing stupid amounts of RAM/swap.

I think that relaxing filtering and sampling and maybe skooching up features a bit will likely get you where you need. If not, maybe increasing the matcher neighbors, especially for the pillars’ sake.

1 Like

I can help you with setting up matcher distance and neighbours. I just need to know front/side overlap and some other values.

I’m in bed now so I’ll follow this up tomorrow!

1 Like

May I ask what the goal is, what is the purpose of the work you are doing?

Some flaws will never go away no matter what settings you use.

1 Like

This has been a great discussion, but all these questions don’t get at the fundamental issue: ODM tends to somewhat aggressively remove tall tree like structures (probably) in the outlier removal routine. That’s where things should be tweaked first.

Pillars match this description… . So look to the point cloud filtering steps for ways forward on this one. :slight_smile:

1 Like

Also other thin linear objects like railings tend to be partially or totally removed, such as the partially missing 40mm X 40mm steel railing here

part-missing-railing

2 Likes

Yup! Probably the same filter.

2 Likes

What does one standard deviation represent?

1 Like

I need:

Front overlap
Side overlap
GSD
Pixel hight
Pixel width

1 Like

https://docs.opendronemap.org/arguments/pc-filter/

Horror. I think I’ve been looking at this parameter backwards -_-

Larger value should be less filtering, so more likely to keep possibly spurious points from reconstruction, but less likely to over-filter.

2 Likes