[Zador-Pataki]

Motivation

Support monocular reconstruction from uncalibrated distorted images while retaining their original image coordinates for matching and reconstruction.

Changes

• Calibration

  • Support OPENCV_FISHEYE and distorted perspective camera models.
  • Use GeoCalib to estimate the initial focal length and distortion needed for tracking and keyframing.
  • After matching and geometric verification, refine focal length and distortion through view-graph calibration before DA3 depth prediction, keeping the principal point fixed.
  • Reuse the refined cameras during mapping without repeating calibration; retain the original GeoCalib predictions as priors.
  • Require initial view-graph calibration whenever intrinsics are estimated, independently of subsequent mapping refinement.

• Tracking and depth

  • Track and match features in the original distorted images.
  • Use the refined calibration to define an auxiliary forward-facing pinhole crop for DA3, then transfer the predicted depths back to the original image coordinates.
  • Keep observations outside that crop usable without depth constraints.

• Reconstruction and visualization

  • Account for the selected distortion model in geometric verification and relative-pose estimation.
  • Extend browser camera visualization and point-quality calculations to the supported distortion models.

Limitations

• DA3 does not receive the original distorted images. It receives a pinhole crop undistorted using estimated calibration, and its predicted depths are resampled back to the original image coordinates. Depth quality therefore depends on that calibration and resampling; depth support is limited to the crop.
• Track sampling, density estimation, and suppression use image-space distances and neighborhoods without compensating for lens distortion. Equal-sized pixel neighborhoods can therefore cover different angular regions across the image, affecting the spatial distribution of tracks.