Forum

Author Topic: Accuracy of masking from low/med quality model  (Read 7332 times)

Andrew

  • Jr. Member
  • **
  • Posts: 77
    • View Profile
Accuracy of masking from low/med quality model
« on: March 31, 2014, 01:14:36 PM »
I read recommendations of masking photos from model on more than one occasion on this forum, but whenever I tried the workflow (align unmasked photos first, build low/med quality mesh and use it to generate masks for photos), there would be enough imprecisions in the mask to scare me away from using this feature. Mask encompassing some of the background is not the end of the world, but often mask would clip tip of the nose, or edge of ears etc.

And so, my questions are:

- where do these imprecisions in masking from model come from? Could it be that masks are generated for undistorted photos but the resulting mask is displayed over original photo that still carries optical distortions? In this case, mask could actually be perfect even if it appears to be off.

- does Photoscan respect photo masks down to pixel precision? Or is there some threshold that causes PS to scan for points even if they fall just slightly behind mask edge?

Cheers,
Andrew



admir

  • Jr. Member
  • **
  • Posts: 93
    • View Profile
Re: Accuracy of masking from low/med quality model
« Reply #1 on: March 31, 2014, 02:07:53 PM »
Hi Andrew, can you detail what are your subject and how are you shooting (single camera, or multicam setup) ? When shooting environment, i only mask out sky if visible, build dense point cloud and delete unwanted geometry before building mesh.
When shooting characters, using empty background to mask out character works very well. I have noticed that you dont have to worry if in some photos you clipped parts of subject if its properly masked in other photos, i.e. if there is enough overlapping. Also, you can fix masks after you aligned cameras anyway and recalc if you are missing some parts you masked by mistake.

Admir

Andrew

  • Jr. Member
  • **
  • Posts: 77
    • View Profile
Re: Accuracy of masking from low/med quality model
« Reply #2 on: March 31, 2014, 11:04:35 PM »
I often shoot in scenarios where hand masking would be too tedious (for instance building rooftop against sky, with occasional tree branches obstructing view). When shooting huge projects, I cant really afford excessive overlap (I risk running out of RAM) so I really need to rely on all available photo information, including what mask from model sometimes clips...

I am curious to hear what Alexey can tell us about causes of mask clipping and whether that is a problem at all :)

-Andrew

chadfx

  • Full Member
  • ***
  • Posts: 205
    • View Profile
Re: Accuracy of masking from low/med quality model
« Reply #3 on: March 31, 2014, 11:46:20 PM »
I have the same question...exactly how close does Photoscan respect those masks and what is the impact of imperfection in them? How strictly are they followed at each step of the pipeline?

It would be great to have a simple grow/shrink mask option as well...just a very basic pixel operation to push out or pull in the pixels, which might help easily trim out some garbage or expand out to encompass some clipped off areas.

Having to manually tweak each individual mask is extremely tedious, and the interactivity of the masking tools is numbingly slow.

-C

FLuca

  • Jr. Member
  • **
  • Posts: 57
    • View Profile
    • Another-me
Re: Accuracy of masking from low/med quality model
« Reply #4 on: April 01, 2014, 03:45:07 PM »
I don't know how exactly the masks are used by PS, however I have another practical question:

Is there any way to remove the masks ??

I need this function when I generate the masks from the model to perform alignement on model only (especially for aligning chunks). As these masks are really not perfectly fitted to the model, I would like to remove the masks to generate the point cloud. Unfortunately I did not find the function and you can not deactivate the mask in point cloud generation (maybe a feature to add in the 'Build Dense Cloud Panel'?).
Fred Lucazeau - Body Scan - www.another-me.fr
Sketchfab gallery: https://sketchfab.com/fredlucazeau

Andrew

  • Jr. Member
  • **
  • Posts: 77
    • View Profile
Re: Accuracy of masking from low/med quality model
« Reply #5 on: April 01, 2014, 09:07:46 PM »
FLuca, to remove masks, right click on your chunk in Workspace pane, and click Remove/Remove Masks

chadfx

  • Full Member
  • ***
  • Posts: 205
    • View Profile
Re: Accuracy of masking from low/med quality model
« Reply #6 on: April 02, 2014, 01:38:56 AM »
...and it would be nice to be able to enable/disable masks instead of having to remove them

FLuca

  • Jr. Member
  • **
  • Posts: 57
    • View Profile
    • Another-me
Re: Accuracy of masking from low/med quality model
« Reply #7 on: April 02, 2014, 03:55:33 PM »
thank you Andrew for the tip !
Fred Lucazeau - Body Scan - www.another-me.fr
Sketchfab gallery: https://sketchfab.com/fredlucazeau

Andrew

  • Jr. Member
  • **
  • Posts: 77
    • View Profile
Re: Accuracy of masking from low/med quality model
« Reply #8 on: April 05, 2014, 10:33:05 AM »

- where do these imprecisions in masking from model come from? Could it be that masks are generated for undistorted photos but the resulting mask is displayed over original photo that still carries optical distortions? In this case, mask could actually be perfect even if it appears to be off.

- does Photoscan respect photo masks down to pixel precision? Or is there some threshold that causes PS to scan for points even if they fall just slightly behind mask edge?


Alexey, could it be possible for you to shed some light in this matter? Masking from model seems like a very useful feature, but for now, it seems nobody fully understands how it works (judging by lack of answers).

-Andrew

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14796
    • View Profile
Re: Accuracy of masking from low/med quality model
« Reply #9 on: April 14, 2014, 02:39:33 PM »
Hello Andrew,

During mask from model generation all distortions are taken into account, so the imprecision you are speaking about is likely to be caused by two factors:
- inaccurate/low quality model,
- inaccurate camera alignment results.
There could be several ways of model based masks improvement, starting from manual mask refinement up to several cycles of alignment, model generation and mask creation.

Masks in PhotoScan are pixel accurate and all the pixels that are covered with masks will not be used in processing, except Align Photos stage with unchecked "Constrain features by masks" option.
Best regards,
Alexey Pasumansky,
Agisoft LLC