Getting small distorted/blurry areas on scan and question about depth filtering.

[loggie24]

Hi there.

I have been scanning quite a lot of rocks lately and i can't help to notice something that happens on nearly all the scans i do. The texture is has this distorted and blurry look in some areas. While it isn't noticeable from a distance when you get up close it really annoys me. I always shoot with a tripod at a fairly high shutter speed so camera shake should not be a problem, but i'm rather curious if it can be something with the settings i'v been using. Whenever i estimate image quality i usually get numbers between 0.7-0.9 which is technically very acceptable.

So my second question is about depth filtering and what setting is the most useful for me. I have red a lot of the documentation and all i'v been able to learn about it is that the "aggressiver" depth filtering the more points/detail (?) will be removed (things that might be a bit out of focus etc), is this correct?

My most used settings are:

Align photos: accuracy - highest, pair preselection - disabled, key point limit - 100000, tie point limit - 50000
Optimize cameras: default
Build dense cloud: quality - high, depth filtering - disabled (this might be a really good or really bad setting)
Build mesh: pretty much default
Build texture: default

Here is a picture of my annoying little problem:



Any help would be appreciated.

[frank.stremke]

this sometimes hapens with stuf from the back or foreground of the focus point being used for texturing you can fight this by agressive masking and trying to get as much stuf in focus as posible (depth of field, apperature)
frank

[Alexey Pasumansky]

Hello loggie24,

Can you post also the screenshots of the dense cloud and mesh for the same area?

[JMR]

Hello Loggie:
IMHO If you are doing things like the case you showed, then you can set depth filtering as agressive. It will remove points only if they're alone in the air that is to say points very distant from their neighbourhood. You might want to disable such filtering if you feel it is removing small or thin details you want to have in the model (not likely to happen in a rock).
If your photos are all sharp and clean, then the only good reason for having blurry areas is local missalignment or bad geometry, so the first question it is: Is the shape (mesh) actually an accurate copy of the object?
1 check alignment error for the individual photos (in the right utmost part of the table in the reference panel, usually hidden to user). they should be under 0.2 pixels, and if not, you can refine the sparse cloud by killing points with reprojection errors higher than 0.5 pix.
2 Are you using autofocus with your camera? if so, you are exposed to problems at this distance if Photoscan is trying to use a unique camera model for all photos. Some lenses exhibit an effect called focus breathing that can prevent good results even using "fixed" lenses unless each photo is calibrated as a diferent camera (try split lens group and then optimize)
3 Holes or gaps in the dense cloud can be closed during the mesh creation stage, but this artificially created areas do not (or might not) correspond to object's geometry but still can be textured if are shown in a single photo... sometimes they look badly textured.
I wish I have helped to find the cause of your problem. I think this rock must give a perfect output with no excuses.

Regards,
GEOBIT

[loggie24]

Thanks for the replies so far.

Frank, i'm currently shooting at f/11 on my Sigma 35mm f/1.4 lens so everything should be fairly well in focus.

Alexey, here you go. As you can see, the mesh is fine and while it might simply be the low res look of the dense cloud it seems not to have much issues.



JMR, thanks for the extensive reply. A lot of the rocks that i will scan in the near future will most likely have moss on them, which probably means i want to disable depth filtering. My question is how do i check alignment errors?  I am using autofocus, could you explain a bit more what you mean by splitting lens groups? I am using a fixed 35mm lens.

[JMR]

To check alignment individual errors open the reference panel and scroll to the right (I'm not sure if photoscan standard gives such info, are you using standard?)

https://vimeo.com/56969614 this video will show you what lens breathing is.
If you are using high f numbers you may think that a bigh depth of field solves most of your problems. It is true but focusing or autofocusing actually changes internal geometry (field of view) and is somehow the same as zooming between shots. This is BAD.
Photoscan will think always you are using one camera with a truely fixed lens and you are not. So if you go to Camera calibration dialog, you will see one single camera/lens that is being modeled by a unique set of parameters.
You can right click on the camera name and split so every photo is considered a diferent camera/lens combo and optimized separately.

Moss wont be removed so easily by depth filtering, so I still would recommend to use it agressively

[loggie24]

Thanks a lot. I checked the alignment errors and no one were under 0.2, mostly around 0.12 so i take it as a good sign. As for the camera breathing problem, i have gone ahead and done as you told me to split group the individual photos. Will re-process everything and see what difference this makes.

Just for the record, could you or anyone explain in which scenario it would be useful to use aggressive/disabled depth filtering? Different objects for example.

[Alexey Pasumansky]

Personally, I don't recommend to use Disabled depth filtering in production, but only for some experiments and research. Aggressive is default option and usually is applicable to well-textured objects/surfaces. Using Moderate/Mild options may be reasonable for the surfaces that do not have much texture details where low number of matching points can be found (for example untextured roofs in aerial survey may be reconstructed with more sharp and accurate edges using Mild option, since the Aggressive may just filter the related points as noise).

[JMR]

I always have found that not filtering gives much more noise in the cloud while aggressive (kinda strong name) is actually the best setting and it does not mean washed or weared models. In addition filtering can remove thin trees or branches, weird point "trails" to the air in reflections and underground or flying points caused by moving objects... honestly I almost never disable aggressive filtering and I'm happy with the results
best

[loggie24]

After re-processing everything i couldn't see any difference, but after a quick look through the pictures i found out that at  least on this specimen i didn't take good enough pictures from the top, most of the were on the side and angled.

Anyways, thanks a lot for all the responses, learned a couple new tricks that i bet will be useful in the future.