good photos, good dense cloud, BAD TEXTURE

[majou]

Dear all,

I am facing the weirdest problem and I am all out of ideas so I really wish someone could help.

Photogrammetry documentation is an essential part of my work, and having a good textured model is one of the desired outcomes that I seek.
just recently I had been having a very poor textured model for every project. it is weird because I am still using the same equipment (camera, ..etc) and using same processing settings in Metashape as I used to do before.
before I used to have an almost photo-realistic textured model outcome but now all of a sudden every project I process is turning out very blurry mushy like texture. funny enough the dense cloud and shaded model looks better than the textured one!

I have tried updating , uninstalling, and reinstalling the software but with no luck.

Can some one please help, any ideas are appreciated. I had been looking for a solution for quiet some time now and I really need one because it affects my work.

thanks!

attached are some explanatory images.  you can clearly see the result I used to have previously (image C)

[Arie]

Hi majou,
could you show the processing settings for the recent and the old project (right-click on chunk, show info)?

[majou]

Dear Arie,

thank you for your reply. attached are the screenshots. project A and B are recent and of which I encountered the problem. project C and D are older (processed about a month ago) and their result was satisfying.

also, I tried processing project C again from zero in a complete new file, the programe crashes on the "building texture" step.

[Arie]

Hi majou,
thank you for the screenshots. Since all settings regarding texturing are identical (except c being processed with 1.63), the error should be somewhere else.
There are some settings, which you can optimize and those might influence the quality of your model including the texture:

1. For alignment, please try accuracy "High" so that feature detection will be computed on all pixels. With "medium" setting, you lower the resolution, which can lead to a slightly larger error for the camera pose.

2. Are you familiar with the optimizing process? It can improve the quality to delete stray points in the sparse cloud including points of the surrounding area of the trench (like vegetation and other "moving" subjects) as well as points which were detected in the out-of-focus areas and to run a camera optimization (the star symbol under "Reference"). You can also run the gradual selection to automatically remove points with a high reprojection error (under Model - Graduatial Selection...).

As an example: When I process excavation sites, I first run the alignment (Accuracy: High). After creating the sparse cloud, I add my GCPs to get an inital impression of the accuracy. I uncheck all markers (so they are used only as checkpoints) and start cropping the sparse cloud i.e. removing the surrounding area (usually full of shrubs, grass etc. which move ever so slightly during image acquisition). Additionally, I run the gradual selection with a reprojection error of about 0.25 - 0.3 (watch the number of sparse points selected) and delete those points as well. After that I run the "Optimize Cameras" using "adaptive camera model fitting". Check the accuracy of your GCPs and, if necessary, use some for adjusting the camera alignment (some should always stay unchecked as validation).

3. You should try using Agisofts new algorithm for computing the high-res model: Instead of building the "dense cloud" go directly to "Build mesh" and use "Source data: depth maps". Please note, that if you reprocess old projects, the depth maps from the dense cloud might still be present- so make sure to uncheck Reuse depth maps (under Build Mesh - Advanced settings).* The quality and speed of this approach is a significant improvment over the old algorithm.

4. When texturing your model you should not hesitate to use texture sizes larger than 4096. Of course it depends a lot on your further use but I frequently use textures sizes from 8192 to 16384 px and sometimes even more. Performance-wise it is better to have one larger texture than multiple smaller ones (e.g. one 8192 x 8192px texture equals four 4096 x 4096px textures).

Good luck!

*Sorry, bad advice on my behalf.

[dpitman]

You can also run the gradual selection to automatically remove points with a high reprojection error (under Model - Graduatial Selection...).

... Additionally, I run the gradual selection with a reprojection error of about 0.25 - 0.3 (watch the number of sparse points selected) and delete those points as well.

3. You should try using Agisofts new algorithm for computing the high-res model: Instead of building the "dense cloud" go directly to "Build mesh" and use "Source data: depth maps".

Hi Arie,

As a noob to Metashape, I'm always on the lookout for nuggets of workflow knowledge.  The manual is pretty basic and I don't know where else to look to better understand the software. So thanks for taking the time to help out majou.

Regarding your comments about removing points from the sparse cloud that fall outside of the threshold you set, how does that effect the reconstruction of the mesh, or dense cloud?  As I understand it, the processor doesn't use the sparse cloud to build the other products so I'm confused on how removing low quality points from it helps in the end?

Regarding going straight to mesh and skipping dense cloud construction, does that mean that you do not do any classification or manual removal of objects  that otherwise will be in your mesh?  For me, removing man-made objects and removable vegetation is important in getting to a "clean" surface model.  How do you do that, if possible, going straight to mesh?

Thanks again for the contributions!

Dave

[Arie]

Hi Dave,
i'll try to keep it short and simple. Regading your first question, the sparse cloud represents feature points, which were used to calculate the camera pose*. If you delete points from the sparse cloud either by manually selecting or using the gradual selection tool and you use "Optimize cameras" the camera pose will be recalculated without the deleted points (warning: if you delete too many points, it might not be possible to re-align the images).
This can help to increase the accuracy of the camera pose since some feature points might be "unstable". For example, vegetation can move slightly during the image acquisition. If these points are used, the overall accuracy of the camera pose decreases. And the more accurate your camera pose, the better quality your high-resolution model will have.

With the new depth-map based approach to calculating the high-res model you will have to clean-up the mesh instead of the pointcloud. But it is a lot more efficient, so you do not have clean as many outliers etc. as with the dense cloud approach. AFAIK you cannot classify the mesh but it should be possible to export the mesh and use Cloudcompare to extract the points of the mesh and reimport those to Agisoft for classification. Bu I haven't tried that yet.
The quality and speed of the depth-map based approach is quite an improvement and I would highly recommend it.

Hope that helps.
Cheers!


* just in case you are not familiar with the term, camera pose describes the position and orientation of the images (i.e. the result of "Align images").

[dpitman]

Good stuff, Arie.  Thanks!

[Costas]

3. You should try using Agisofts new algorithm for computing the high-res model: Instead of building the "dense cloud" go directly to "Build mesh" and use "Source data: depth maps". Please note, that if you reprocess old projects, the depth maps from the dense cloud might still be present- so make sure to uncheck Reuse depth maps (under Build Mesh - Advanced settings). The quality and speed of this approach is a significant improvment over the old algorithm.

Hello Arie,

regarding step 3.

I'm a bit confused on what you said because there are cases that we need to calculate "Dense Cloud" and also have a "Mesh". My question has to do with the "Reuse Depth Maps".

Lets say that i have "Dense Cloud" with High Quality Depth Maps and i now want to build a "Mesh". If i want to use the same quality (High) Depth Maps which are calculated before why not using "Reuse Depth maps"? Are the depth maps calculated in Dense cloud different to the ones calculated in Mesh or you said "do not reuse depth maps" because with the new algorithms you can do even better quality (eg. Ultra) depth maps with the same PC resources, thus better Mesh?

The same applies to old projects that we already have the Depth Maps. If we do not want to change the Quality why not build a Mesh selecting the "source" as Depth Maps and "Reuse Depth maps". Will it not result in the same quality?

[Arie]

Hi Costas,
since I don't reprocess old projects super frequently, I might have made an error. I did a little research and in an older thread I found this statement from Alexey:

"If the depth maps are present in the project and you are selecting the further processing option that assumes depth maps generation, then you will be able to re-use them (skip there generation again), providing that the same quality/filtering settings are selected.
Note that new mesh currently requires only depth maps generated with Mild filtering option, so in your workflow you would be able to re-use depth on Step 6 (build mesh) only if on set 3 (build dense cloud) you have used Mild filtering option, and on Step 6 select the same quality as on the Step 3."

The last sentence seems to be not true anymore. I just tested it with depthmaps generated with "mild" and "agressive" filters (dense cloud) and both could be reused in the meshing step. In my quick comparision, the depthmaps (and resulting mesh) with "Mild" filter settings (dense cloud) are equal to the new depth-maps created with "Build mesh". In "Build mesh" you cannot adjust the filter anymore, its automatically set to "mild" when calculating new depthmaps.

I probably misinterpreted my results when testing the two approaches and actually compared mild/ agressive filter settings and not the actual resulting depthmaps.
Thanks for pointing that out!

[majou]

for any one stopping by this thread, facing similar issue, thanks to Arie and his advice I had been able to have good texture back.
Somehow, the result of poor blurry textured model is a sum of accumulative poor decisions. by choosing "medium" quality for the alignment and the dense cloud, and later using small tile size for the texture.
so, the better practice leading to good result is: running alignment on "high" quality, using "optimize photos" , processing mesh while skipping the dense cloud step (unless needed) on high quality, and finally using a larger texture tile size for processing texture.

hope you all find this useful as I did.

[Steve003]

Hi,
is this metashape or Photoscan and what version ?
where is the optimize photos stage ?
metashape I use on 'High', generic preselection ticked, estimated, key point 40000 tie point 10000, exclude stationary tie points. guided image matching off adaptive camera off.

then mesh from depth maps. qlty high face count high, interp enabled, calculate vertex colours and reuse depth maps ticked.

and I am getting blurred areas on textures when make texture is run., whilst the iamges prior to that look fine and far better. There are no blurs in the original images.

This is a major problem, yet Photoscan did same project and no blurs.
Using Photoscan instead though sees what took 1 hour taking 5hrs so far.
Has metashape got a weakness here ?

update:-
and I have just read in a different part of the forum , bug reports

https://www.agisoft.com/forum/index.php?topic=13701.0

Its a bug in 1.7.4

Apparently 1.7.1 is without the bug.

Steve