Author Topic: Camera calibration distorts reconstructed model  (Read 279 times)


  • Newbie
  • *
  • Posts: 1
    • View Profile
Camera calibration distorts reconstructed model
« on: August 28, 2020, 11:14:06 AM »
Hello all,

I'm working as an accident investigator. We use GoPro fisheye cameras to completely scan accident sites and reconstruct the 3D model with MetaShape.

This works fine, when the accident site has a compact size. But we have been stuck on the reconstruction of a site, that ist approximately 150 m long and only 10 m in width. The site has been recorded with 1266 pictures, all oriented in the same direction to not confuse the SIFT algorithm or the camera alignment. During recording of the site, the camera was held pointing down to the surface of the street. The generated report of the project says, every cameraposition overlaps with at least 9 other cameras.

For camera alignment we obviously made sure, that MetaShape uses the fisheye camera model.

Although we made sure to scan the site with 4 rows of overlapping (to the side and to the top and the bottom of the picture) pictures, simultaneously recording the GPS Data of the camera positions, the camera alignment proces keeps reconstructing the accident site in sort of a bent shape. This happens although the GPS data - as a starting point for the alignment process - shows, that the accident site is completely flat in reality.

The reason seems to be that the reconstructed camera calibration deviates just a little bit from the correct camera calibration, which causes the reconstructed model to be bent over its whole length (see attached picture of the sparse point cloud).

We used Markers (spread over the accident site) to support the alignment process as well, but is there a possibility to just assign a height/Altitude parameter to a detected marker without assigning the other parameters like longitude and latitude?
Or do we need to introduce our own coordinate system on site, so that we can place markers to define our own coordinate system to scale the model and to introduce a reference for alignment?

How can we influence the alignment process in a way, that the camera calibration is determined correctly?

Thanks for your help!

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 11878
    • View Profile
Re: Camera calibration distorts reconstructed model
« Reply #1 on: September 04, 2020, 11:20:45 PM »
Hello Winkelschraube,

If you have the control points with the precisely known altitude, you can set approximate Lat/Long values for them and assign low accuracy in XY plane and high accuracy along Z (Set Accuracy: 20/0.1, for example, which means 20 meters in XY and 10 cm along Z). and use for the alignment optimization.

Also you can adjust camera accuracy, if the coordinates are measured more accurately than 10 m.

Please also check that there are no misaligned cameras which may form those rays of tie points that are shown on the screenshot.
Best regards,
Alexey Pasumansky,
AgiSoft LLC