Authors: Alexandra I. Papadaki, Ronny Hänsch Description: Keypoints that do not meet the needs of a given application are a very common accuracy and efficiency bottleneck in many computer vision tasks, including keypoint matching and 3D reconstruction. Many computer vision and machine learning methods have dealt with this issue, trying to improve keypoint detection or the matching process. We introduce an algorithm that filters detected keypoints before the matching is even attempted, by predicting the probability of each point to be successfully matched. This is realized using a flexible and time efficient Random Forest classifier. Experiments on stereo and multi-view datasets of building facades show that the proposed method decreases the computational cost of a subsequent keypoint matching and 3D reconstruction, by correctly filtering 50% of the points that wouldn’t be matched while preserving 73% of the matchable keypoints. This enables a subsequent processing with minimal mismatches, provides reliable matches, and point clouds. The presented filtering leads to an improved 3D reconstruction of the scene, even in the hard case of repetitive patterns and vegetation.