Assessment of predictive clustering trees on 2D-image-based Ear recognition

Abstract

In the last decade person recognition based on various biometric metrics have steadily been gaining on popularity. The same holds for machine learning approaches and various image classification and retrieval techniques. However, many techniques rely on distinguishing between significantly dissimilar images, which is often not the case in person recognition. Person recognition based on images relies on detecting minor differences and not global appearance of an image. To test if retrieval approaches based on bag-of-words fail in the task of biometric recognition we evaluated the following procedure. Ear images were used to extract Scale Invariant Feature Transform feature vectors. These vectors were then fed into forest of Predictive Clustering Trees, k-means and approximate kmeans; and then compared to baseline system where only distances between plain descriptors are compared. While these methods have been proven to perform well in image with significantly different content, the results show that these methods do not perform well under the task of ear recognition.