We propose KRAFT, a computationally efficient feature matching algorithm for small-baseline stereo that generates dense, high-precision correspondences using raw patch descriptors and an optional geometry- and structure-aware filtering cascade, achieving state-of-the-art precision (98.9%) and outperforming standard methods on datasets including KITTI, Middlebury, and human-centric scenes.
KRAFT: Keypoint Robust and Adaptive Feature Tracking
An efficient feature matching algorithm for small-baseline stereo that produces dense, high-precision correspondences using raw patch descriptors and a geometry-aware filtering cascade, achieving state-of-the-art precision.
Short: KRAFT
Projects involved: