On-the-job self-calibration of lens distortion with a Brown model

PCdurham · 18 May 2018 06:18

Congrats to all who are making WebODM and ODM a reality. It’s now by far the most effective and easy to use open SfM package. The WebODM interface is georgeous and it’s getting close to rivalling the big commercial packages: Photoscan and Pix4D. I know a lot of academic end-users (think geographers and archaeologists) who would prefer to do research with open source software. But the fact that you need to go through the checkerboard approach to calibrate lens distortion is a significant drawback. From an academic perspective, there is no evidence that these little drone cameras are very stable over time, meaning that you should redo the calibration shoot. And there is also limited evidence in the academic literature that solutions perform better and have less surface noise without a fixed calibration. From a business perspective, this is the sort of little extra step that would push a business towards the commercial solutions (or lead them to make distorted models!). Ultimately there is a reason the big commercial guys (and MicMac as well) have moved away from the checkerboard approach. I can imagine that there are computation implications. But having a Brown-Conrady lens distortion built into the self-calibration would bring WebODM/ODM in the league of major players and eliminate the last remaining substantial functionality difference. Please consider it.

Patrice

smathermather · 20 May 2018 02:43

Hi Patrice,

Thanks for the heads-up. Can you add an issue on the OpenDroneMap/OpenDroneMap repo? It looks like the underlying work is there in OpenSfM:

github.com

mapillary/OpenSfM/blob/7e393135826d3c0a7aa08d40f2ccd25f31160281/opensfm/src/bundle.h#L305


      
            double scale_;
          };
          
          
template <typename T>
          void BrownPerspectiveProject(const T* const camera,
                                       const T point[3],
                                       T projection[2]) {
            T xp = point[0] / point[2];
            T yp = point[1] / point[2];
          
          
  // Apply Brown--Conrady radial and tangential distortion
            const T& k1 = camera[BA_BROWN_CAMERA_K1];
            const T& k2 = camera[BA_BROWN_CAMERA_K2];
            const T& p1 = camera[BA_BROWN_CAMERA_P1];
            const T& p2 = camera[BA_BROWN_CAMERA_P2];
            const T& k3 = camera[BA_BROWN_CAMERA_K3];
            T r2 = xp * xp + yp * yp;
            T radial_distortion = T(1.0) + r2  * (k1 + r2 * (k2 + r2 * k3));
          
          
  T x_tangential_distortion = T(2.0) * p1 * xp * yp + p2 * (r2 + T(2.0) * xp * xp);
            T x_distorted = xp * radial_distortion + x_tangential_distortion;

Pull requests also welcome , and we’re happy to help you prepare it.

Cheers!

smathermather · 12 December 2019 03:30

Thanks for this advice, Patrice. You were so right: Camera types, a small survey - #2