Author: Schonerger
Year: 2016
- General open source framework for SfM COLMAP
- SfM for unordered image set
- Incrementally reconstruct the scene starting from a two view reconstruction
- Build a graph of overlapping images with kp and descr
- Perform geometric verification with
$E$ (for calibrated cams) or$F$ (for uncalibrated cams) to obtain a verified scene graph with verified image pairs - A good image pair selection is necessary for good reconstruction: better if it is a scene densely populated in images
- New images are added by solving PnP (with or without intrinsics in pb)
- Triangulate points for each new image too
- BA is needed to prevent drift to an unrecuperable state
- Performs local BA at each update on the set of most connected images, performs global BA only after growing largely the model
- contributions:
- better geometric check: do both homography and epipolar and compare the number of outliers
- better best view selection: Use a pyramidal grid approach to compare overlapping between image
- better multi view triangulation: formulate multi view triangulation with RANSAC
- Iterative BA with retriangulation and outlier filtering
- BA: sparse direct solver for few hundreds of cams, and preconditionned conjugate gradient for more, remove cams and observations leading to big errors, re triangulate (RT) after BA, perform iteratively {BA, RT, filtering} until no outliers are filtered
- Deal with redundant view: group highly overlapping images and parametrize it as a single camera
- UNCLEAR: how it deals with internal camera parameters