Historical Image Processing - Can't get GCP/check RMS down


I’m processing some B&W historical imagery from the 30’s, and can’t get the horizontal RMS of many of my check points under 10m in specific areas of the project. There is limited 3 photo overlap, no camera model. I’ve attached my report and the location of my tie points (3 image overlap) andarrows of major issue areas. I understand how my accuracy would degrade outside of the boundary of the control, but can’t win the tug of war with control inside the area. I feel like the likely culprit is matching issues of the photos in those areas as the error is primarily in the X and the flights run E-W. Be interested in your ideas…some questions of mine are:

  • How important is an accurate image center (camera position) location to the matching process? Mine are about 20m horizontally and ~100m vertically as I have to estimate them.
  • I’ve only tried to place GCPs in areas with 3 or more image overlap. Due to the uncertainty in my source data should I only use GCPs in even higher overlap?
  • What would be possible causes of a systematic horizontal bias? I believe the solution is 30m not wide enough for the terrain.
  • How can I densify the matches in the south end? There aren’t many features of note there.
  • Would adding a lot of manual tie points in the seams of three image overlap help refine the solution better? Or other?

Thanks for thinking about this challenge.

1 Like

1 Like

Hi Steven,

What an interesting project :slight_smile:

To answer your questions, here are our first suggestions:

  • The software can process images without geolocation. It is better to rely on GCPs only when you don’t have geotags on your images, and not to estimate them. Reconstruction is a very finely-tuned process and estimating the locations would only mislead it.
    To refine the georeferencing, you could add scale and orientation constraints, but it might be difficult to find objects you can rely on in your pictures, given their age.
  • It is best to mark GCPs in 3 images minimum. For example, we would not recommend to use GPC 40, which you could mark in only two images. Same goes for your check points.
    Also, mark them as precisely as possible, as it will also contribute to lowering the reprojection error (should be around 1.0 at worst).
  • The horizontal bias issue should be solved once you do not estimate the images’ georeference anymore. As the software gives more weight to GCPs than image geolocation, and since you estimated the latter, the horizontal error will be substantial.
  • To densify the matches in the south end, you could try to change the Image Scale parameter (article).
  • It is always good practice to use MPTs, as it helps reconstruction. We usually recommend adding 5-10 MTPs in areas where matching is difficult, but sometimes more are needed.


Hope this helps!

1 Like