Forum

Show Posts

This section allows you to view all posts made by this member. Note that you can only see posts made in areas you currently have access to.


Messages - Vlad

Pages: [1] 2
1
General / Texture reprojection subsampling settings.
« on: December 08, 2022, 02:36:56 PM »
Hello.

I found that in texture reprojection Agisoft does a zero subsampling by default.
That makes this operation looks insane fast compared to other tools. But quality wise this just introduces ugly aliasing.

Did we have somewhere in Advanced options or Tweaks any way to define how many subsampling should be used?
4/16/64 subsamples should be enough to match projection quality with others.

Thank you.

2
Bug Reports / Re: Checker grid like lines in depth map generation.
« on: January 20, 2020, 04:44:23 AM »
can confirm bug.
looks like depth map used for mesh are 8bit.
and this bug from beta versions.

3
General / Re: Alignment Experiments
« on: April 03, 2019, 02:46:57 PM »
Original post lost all images, due to picture hosting provider.
So i decided to make a copy of original post, because such valuable information should not lost in time
https://medium.com/@ssh4/alignment-experiments-from-agisoft-forum-by-marcel-19b2c7727d7a

4
ok, file sent.

5
I have two problems with importing cameras and undistorted images exported from RealityCapture.

120 cameras from camera rig. Alignment report in RC 0.5px reprojection error, mean and median reprojection errors about 0.2~0.3px.

Export cameras as Bundler .out v0.3 z-inverted. and undistorted images.

Photoscan import images and bundler without problem, but when i try check reprojection error Gradual selection show me errors around 3000-4000px. And texturing exported from RC meshes create awful result.

For check issue i export and import bundler from RC back to RC. And report have mostly the same sizes as in original RC project.
But if i export imported to Photoscan cameras as Bundler and open this file in RC it show me same as in Photoscan wrong dimension: 4000-3000 px error.

Cameras in project have different lens settings.



And second issue that i found workaround for this moment, but it still weird. If project have any cameras with vertical orientation export this camera as Bundler and import this bundler to Photoscan show wrong orientation (all cameras are horizontal).
Workaround it clear rotation tag and use in project only horizontal images. In this case imported bundler project show cameras in correct orientation.


This problems exist in latest stable and beta releases. In Standard and Pro versions. Please fix this issue as soon as possible.

ANd here the files for tests:

RC Original: https://www.dropbox.com/s/vc44zocnf82iatx/20171019-2031-RC_Lab_mm.out?dl=0

RC-> Bundler -> AP -> Bundler: https://www.dropbox.com/s/jaim56afn8gumhc/20171019-2031-RC_Lab_mm_AP.out?dl=0

AP origina (from original images aligned in Photoscan)l: https://www.dropbox.com/s/x71xpz1l5gj9mn9/ME4.out?dl=0

6
General / Re: Agisoft PhotoScan 1.4.0 pre-release
« on: September 22, 2017, 10:59:56 AM »
Interesting how far Photoscan users from care about mesh quality :)
Anything but not most impressive feature of AP 1,4.

Some example how photo-consistency refinement work in AP.
But this result still far away from what this method can do.

https://i.imgur.com/p9PDoHx.png

And this is Medium quality ;) that required 2Gb of VRAM for 53x16Mpx only images (as i understand used 3 or 4 times downsampled images)

7
General / Re: Agisoft PhotoScan 1.4.0 pre-release
« on: September 20, 2017, 04:30:12 PM »
Even if user decimate mesh to a deep **it and will play with 50images, such heado-on solution will always have not enough memory. At least OpenMVS can use CPU and RAM. And 300+ images still possible to process on 64Gb RAM.

Currently refine mesh functionality is experimental feature in PhotoScan and is available on GPU only (CUDA / OpenCL), but we are planning to implement CPU support for this feature in the next updates.

As for the video memory consumption, it depends not on the number of images, but the image resolution and quality option selected in Refine Mesh dialog. Also the mesh size is important, but it's impact is considerably lower than the one related to the images.

Yes i know, i played with this method on OpenMVS, with exactly the same limitations. So CPU version also did not help too much. 64Gb RAM was not enough for 300+ 16Mpx images because all processes going in core.

Do you have any information regarding memory requirements in your implementation?

53x16Mpx images can be processed at High settings, not Ultra.

But what VRAM needed for example for 500x24Mpx on High? Or On lower?

8
General / Re: Agisoft PhotoScan 1.4.0 pre-release
« on: September 20, 2017, 11:44:35 AM »
That's funny. Did you just copy and paste code from OpenMVS RefineMesh module for your mesh refinement?

Even if user decimate mesh to a deep **it and will play with 50images, such heado-on solution will always have not enough memory. At least OpenMVS can use CPU and RAM. And 300+ images still possible to process on 64Gb RAM.

2017-09-20 16:28:32 RefineMesh: quality = Ultra high, iterations = 20, smoothness = 1
2017-09-20 16:28:32 Using device: GeForce GTX 960, 8 compute units, 2048 MB global memory, compute capability 5.2
2017-09-20 16:28:32   max work group size 1024
2017-09-20 16:28:32   max work item sizes [1024, 1024, 64]
2017-09-20 16:28:32 Analyzing mesh detalization...
2017-09-20 16:29:05 Memory required: 1743 Mb + 6 Mb = 1749 Mb
2017-09-20 16:29:05 Stage #1 out of 4
2017-09-20 16:29:05 Subdividing mesh...
2017-09-20 16:29:07 Memory required: 27 Mb + 6 Mb = 33 Mb
2017-09-20 16:29:07 Loading photos...
2017-09-20 16:29:09 loaded photos in 2.642 seconds
2017-09-20 16:29:09 Refining model...
2017-09-20 16:29:12 Iteration #1 out of 20
2017-09-20 16:29:40 Iteration #2 out of 20
2017-09-20 16:29:59 Iteration #3 out of 20
2017-09-20 16:30:07 Iteration #4 out of 20
2017-09-20 16:30:16 Iteration #5 out of 20
2017-09-20 16:30:28 Iteration #6 out of 20
2017-09-20 16:30:49 Iteration #7 out of 20
2017-09-20 16:31:12 Iteration #8 out of 20
2017-09-20 16:31:34 Iteration #9 out of 20
2017-09-20 16:31:54 Iteration #10 out of 20
2017-09-20 16:32:16 Iteration #11 out of 20
2017-09-20 16:32:38 Iteration #12 out of 20
2017-09-20 16:32:56 Iteration #13 out of 20
2017-09-20 16:33:14 Iteration #14 out of 20
2017-09-20 16:33:38 Iteration #15 out of 20
2017-09-20 16:33:57 Iteration #16 out of 20
2017-09-20 16:34:20 Iteration #17 out of 20
2017-09-20 16:34:39 Iteration #18 out of 20
2017-09-20 16:35:00 Iteration #19 out of 20
2017-09-20 16:35:09 Iteration #20 out of 20
2017-09-20 16:35:17 Stage #2 out of 4
2017-09-20 16:35:17 Subdividing mesh...
2017-09-20 16:35:20 Memory required: 108 Mb + 19 Mb = 128 Mb
2017-09-20 16:35:20 Loading photos...
2017-09-20 16:35:23 loaded photos in 2.516 seconds
2017-09-20 16:35:23 Refining model...
2017-09-20 16:35:31 Iteration #1 out of 20
2017-09-20 16:35:58 Iteration #2 out of 20
2017-09-20 16:36:25 Iteration #3 out of 20
2017-09-20 16:37:01 Iteration #4 out of 20
2017-09-20 16:37:42 Iteration #5 out of 20
2017-09-20 16:38:09 Iteration #6 out of 20
2017-09-20 16:38:37 Iteration #7 out of 20
2017-09-20 16:39:05 Iteration #8 out of 20
2017-09-20 16:39:32 Iteration #9 out of 20
2017-09-20 16:39:59 Iteration #10 out of 20
2017-09-20 16:40:26 Iteration #11 out of 20
2017-09-20 16:41:02 Iteration #12 out of 20
2017-09-20 16:41:34 Iteration #13 out of 20
2017-09-20 16:42:08 Iteration #14 out of 20
2017-09-20 16:42:42 Iteration #15 out of 20
2017-09-20 16:43:15 Iteration #16 out of 20
2017-09-20 16:44:33 Iteration #17 out of 20
2017-09-20 16:46:20 Iteration #18 out of 20
2017-09-20 16:46:52 Iteration #19 out of 20
2017-09-20 16:47:25 Iteration #20 out of 20
2017-09-20 16:47:56 Stage #3 out of 4
2017-09-20 16:47:56 Subdividing mesh...
2017-09-20 16:48:07 Memory required: 435 Mb + 72 Mb = 508 Mb
2017-09-20 16:48:07 Loading photos...
2017-09-20 16:48:10 loaded photos in 2.516 seconds
2017-09-20 16:48:10 Refining model...
2017-09-20 16:48:42 Iteration #1 out of 20
2017-09-20 16:50:57 Iteration #2 out of 20
2017-09-20 16:52:44 Iteration #3 out of 20
2017-09-20 17:05:25 Iteration #4 out of 20
2017-09-20 17:07:18 Iteration #5 out of 20
2017-09-20 17:08:58 Iteration #6 out of 20
2017-09-20 17:10:39 Iteration #7 out of 20
2017-09-20 17:12:18 Iteration #8 out of 20
2017-09-20 17:13:58 Iteration #9 out of 20
2017-09-20 17:15:39 Iteration #10 out of 20
2017-09-20 17:17:19 Iteration #11 out of 20
2017-09-20 17:18:59 Iteration #12 out of 20
2017-09-20 17:20:38 Iteration #13 out of 20
2017-09-20 17:22:18 Iteration #14 out of 20
2017-09-20 17:23:57 Iteration #15 out of 20
2017-09-20 17:25:38 Iteration #16 out of 20
2017-09-20 17:27:17 Iteration #17 out of 20
2017-09-20 17:28:57 Iteration #18 out of 20
2017-09-20 17:30:38 Iteration #19 out of 20
2017-09-20 17:32:17 Iteration #20 out of 20
2017-09-20 17:33:57 Stage #4 out of 4
2017-09-20 17:33:57 Subdividing mesh...
2017-09-20 17:34:42 Memory required: 1743 Mb + 287 Mb = 2030 Mb
2017-09-20 17:34:42 Loading photos...
2017-09-20 17:34:45 loaded photos in 3.218 seconds
2017-09-20 17:34:45 Refining model...
2017-09-20 17:36:57 Iteration #1 out of 20
2017-09-20 17:36:58 Finished processing in 4106.3 sec (exit code 0)
2017-09-20 17:36:58 Error: out of memory (2) at line 179

9
General / Re: Agisoft PhotoScan 1.4.0 pre-release
« on: September 20, 2017, 07:42:43 AM »
One biggest problem that Mesh Refinement as usual for Agisoft is not out of core. And 2Gb VRAM not enough for average 2017 years scans.

This is bad, because your users is not rich studios that moved years from Photoscan to more modern tools. But hobbyists that still can have 1Gb or 2Gb VRAM GPU.

>4Mln poly mesh required ~2,3Gb of RAM and do not allow make refinement.

10
General / Re: Agisoft PhotoScan 1.4.0 pre-release
« on: September 20, 2017, 03:36:57 AM »
Why the **ck 1.4 pre even if i define other folder to install removed 1.3 version? :(

But Mesh Refinement is amazing. You finally made this! But because you take one known paper about this method, most of users will did not find any profit comparing to non refined workflow ;)
Mesh refinement required some changes. And can be faster that usual.

11
General / Re: Any significance to dot colors in sparse and dense clouds
« on: September 01, 2016, 05:19:15 AM »
Because Photoscan still use algorithms from early photogrammetry days (that probably as old as mammoth ***t). All Cloud dots are "real" for Photoscan, but depend on mesh generation resolution settings, alone dots or dots islands can be ignored. So you SHOULD remove all garbage from dense cloud before generate mesh.

You can read about poison mesh generation or other technics for understand how this work. Photoscan just use such mesh generation create raw mesh about 1.5~2x more poly than target count and later just decimate for your target poly count.


12
General / Re: bUMPY SURFACE
« on: July 20, 2016, 12:19:33 PM »
This is "weak" surfaces. You need shoot photos of walls perpendicularly +-20degree, more than 45 degree you will gave dof blur. and this will create such blobby dense cloud.

Other  photogrammetry software than photoscan, can create better results with flat surfaces.

13
General / Re: Interior scan question
« on: July 17, 2016, 02:45:25 AM »
Please show camera placement. Probably you just not have enough coverage on walls or walls on photos at angle more than 45ยบ

14
General / Re: about the photo alignment
« on: July 17, 2016, 02:43:17 AM »

I think photogrammetry reconstructions could follow this path in time, simply left to their own devices for as long as desired and the results refine to ever finer details, rather than having arbitrary pre-set limits at the initial camera detection steps etc.

What if I tell you that modern photogrammetry algorithms and tool work like you say. Unfortunately Photoscan still use old way.
If you interesting you can read this paper:
High Accuracy and Visibility-Consistent Dense Multiview Stereo HH. Vu et al. 2012.

I hope Agisoft one day stand from five point and do something to improve quality and speed of it not cheap app.

15
General / Re: about the photo alignment
« on: July 16, 2016, 05:03:32 PM »
May be your photos is not good enough, low resolution that not allow find "features" required for tie point calculation?

Can you share 5-10 photo sequence for understand problem?

Pages: [1] 2