Author: Sattler
Year: 2019
- memory limit for localization in a 3D model
- small set of point is represented with full apearance info but large set of points with compressed info (hybrid)
- CNN method like poseNet
- small set of pt used for high quality correspondences for full descriptor matching
- RANSAC variation to handle both feature types and 3D covisibility
- Averaging SIFT descriptor on all images from which the point was triangulated
- Vocabulary based feature matching, word = one index (integer)
- Scene compression (i.e. reducing the number of points) using covisibility and vocabulary (penalizes words that are over represented)
that selects
$\mathcal{P'}$ the subset of descriptor unique points - big set of point will only be described by word (quantized desc), they will result in multimatch disambiguated by pose
Modified RANSAC:
- Samples with a high probability co visible points in quality match
- P3P for calibrated cameras and P4P when focal is not known
- check inliers with both quality and quantized matches