Keypoints Matching on Uncalibrated Face Stereo Images

Keypoints matching on uncalibrated face stereo reconstruction is generally insufficient due to the sparse facial texture. This paper introduces a novel method based on two-layer processing structure to generate more dense and scattered matching pairs. In the first layer, initial matches are extracted applying an improved LOG detector and an image blocking matching scheme, where the former copes well with sparse texture detection and the latter improves the points scattered distribution. Then based on these initial matching results, RBF is applied to transform all the candidate points between the stereo images to carry out the final matching pairs. Due to the stable facial structure, both the number and accuracy of matches are improved. Experiments demonstrate that the proposed scheme outperforms the other points matching algorithms and generates matching pairs with reliable distribution.