Mirrored non-maximum suppression for accurate object part localization

There has been significant progress in object part localization such as human pose estimation and facial landmark detection. In most of the previous methods, two phenomena are ignored. Firstly, they usually output a set of candidate pose hypotheses but the hypothesis with the highest score obtained by Non-Maximum Suppression (NMS) is not always the optimal result. Secondly, they can not get exactly bilaterally symmetric keypoints on the mirrored images even though the training data is always augmented with mirrored images. In fact, the intrinsic relationship between the original image and the mirrored one is helpful for object part localization. In this paper, we propose Mirrored Non-Maximum Suppression (Mirrored NMS) which can utilize mirrored detections to improve the accuracy of object part localization. Experimental results show that our method can improve the state-of-the-art accuracy by 1.3~3.0% in PCP for human pose estimation and can produce more accurate results than averaging multiple hypotheses for facial landmark detection.