Forum

Author Topic: Accuracy measurement?  (Read 6556 times)

chrisd

  • Full Member
  • ***
  • Posts: 187
    • View Profile
Accuracy measurement?
« on: June 26, 2013, 04:04:16 PM »
Hello,

I searched the manual, forum and the wiki and have some good information about best practices for using the software and camera but did not see anything that answers my question below.

Is there a method that can predict the accuracy for an object (an "approximate" range of values is OK) that could apply to future situations without loading the images into Photoscan to get the values?


 


RalfH

  • Sr. Member
  • ****
  • Posts: 344
    • View Profile
Re: Accuracy measurement?
« Reply #1 on: June 26, 2013, 04:19:29 PM »
It is not possible to make quanitative predictions regarding model accuracy. It depends on many factors, for example sensor resolution, lens quality, object coverage, image overlap, surface texture, illumination, camera-to-object viewing angles etc. Also, depending on the type of object (e.g., full body capture or aerial survey), different image acquisition strategies/setups can be used which further influence model quality.

chrisd

  • Full Member
  • ***
  • Posts: 187
    • View Profile
Re: Accuracy measurement?
« Reply #2 on: June 26, 2013, 07:00:57 PM »
Hi Ralf,

I understand there are MANY variables.

The fewer variables that change, the more predictable the results will be. So assume I use the same camera, lens, distance to subject, take the pictures from the same locations, etc. So essentially the only thing that changes is the subject, but the subject is the same size. Lets also say the subject is static.

Surely the accuracy could be reasonably predicted if the only thing that changes is the subject?

What I would like to find out is if I have a human (or house, or baseball etc.) sized object, all other things being equal, what I could expect for accuracy compared to other similar sized objects.

I am NOT asking for the specific amount of error. I would like to know what a reasonable approach to repeatable accuracy would be. Obviously, you can't put everything in a studio under completely controlled conditions. If I am taking photos of houses, I am not going to take the pictures from the same relative positions every time.

So, if you can't take the pictures from the same relative locations, what would be the next best thing to do, ie same distance?






BrenoGeo

  • Newbie
  • *
  • Posts: 2
    • View Profile
Re: Accuracy measurement?
« Reply #3 on: June 26, 2013, 08:42:13 PM »
Hello Chris,

If you could mark some control points with known three-dimensional coordinates in scenarios that you're photographing, you could use those control points as a basis to guide further work, it may bring more accuracy to their work.

And the distance will not matter.

RalfH

  • Sr. Member
  • ****
  • Posts: 344
    • View Profile
Re: Accuracy measurement?
« Reply #4 on: June 27, 2013, 10:43:59 AM »
All other things (relative distance, relative camera positions, cameras etc.) being equal, model accuracy will depend a lot on object texture. If you have good texture, Photoscan will be able to reliably detect many matching points (hence create a good alignment) and good depth maps (hence low error rate in dense depth maps and final model). If you have poor texture (large blank areas, shiny surfaces such as windows, self-similarity), alignment, depth maps and final model will have poorer geometric quality, sometimes even to the point that alignment becomes impossible. No way to predict this in quantitative terms without processing and analysis of control points for each model.

chrisd

  • Full Member
  • ***
  • Posts: 187
    • View Profile
Re: Accuracy measurement?
« Reply #5 on: June 29, 2013, 04:23:28 AM »
With more experience with Photoscan, I will probably be able to make some "educated guesses" about the accuracy of the results based on the subject characteristics and the shooting conditions.

What I was considering was shooting a bunch of objects in perhaps 3 size ranges, similar to a baseball, a car, and a building. However, based on the feedback so far, this does not sound as useful as I had anticipated to make some generalizations about accuracy.