Abstract: Aiming to quickly and accurately select inliers from initial matching results of local invariant features, an inlier selection algorithm is proposed by jointly exploiting the distribution consistencies of features in location, scale and orientation. Firstly, coarse selection is performed by grouping feature correspondences into clusters according to disparities via the Mean-Shift algorithm. Secondly, K-nearest neighbors of each feature are found in the cluster to which it belongs. Then parameters of transformation between the two local image regions where each pair of features are located are estimated by feature scales and orientations, and a K-nearest neighbor matching similarity is calculated. Finally, according to K-nearest neighbor matching similarities, inliers are selected from the candidate matching set obtained by coarse selection. Experimental results show that the proposed algorithm is superior to the state-of-the-art inlier selection algorithms in terms of precision, recall and speed, and is robust to large changes between images in viewpoint, scale and rotation.