Learning to Find Good Correspondences

摘要: We develop a deep architecture to learn find good correspondences for wide-baseline stereo. Given set of putative sparse matches and the camera intrinsics, we train our network in an end-to-end fashion label as inliers or outliers, while simultaneously using them recover relative pose, encoded by essential matrix. Our is based on multi-layer perceptron operating pixel coordinates rather than directly image, thus simple small. introduce novel normalization technique, called Context Normalization, which allows us process each data point separately imbuing it with global information, also makes invariant order correspondences. experiments multiple challenging datasets demonstrate that method able drastically improve state art little training data.