Human Detection via Classification on Riemannian Manifolds

Abstract

We present a new algorithm to detect humans in still images utilizing covariance matrices as object descriptors. Since these descriptors do not lie on a vector space, well known machine learning techniques are not adequate to learn the classifiers. The space of d-dimensional nonsingular covariance matrices can be represented as a connected Riemannian manifold. We present a novel approach for classifying points lying on a Riemannian manifold by incorporating the a priori information about the geometry of the space. The algorithm is tested on INRIA human database where superior detection rates are observed over the previous approaches. IEEE CVPR 2007This work may not be copied or reproduced in whole or in part for any commercial purpose. Permission to copy in whole or in part without payment of fee is granted for nonprofit educational and research purposes provided that all such whole or partial copies include the following: a notice that such copying is by permission of Mitsubishi Electric Research Laboratories, Inc.; an acknowledgment of the authors and individual contributions to the work; and all applicable portions of the copyright notice. Copying, reproduction, or republishing for any other purpose shall require a license with payment of fee to Mitsubishi Electric Research Laboratories, Inc. All rights reserved. AbstractWe present a new algorithm to detect humans in still images utilizing covariance matrices as object descriptors. Since these descriptors do not lie on a vector space, well known machine learning techniques are not adequate to learn the classifiers. The space of d-dimensional nonsingular covariance matrices can be represented as a connected Riemannian manifold. We present a novel approach for classifying points lying on a Riemannian manifold by incorporating the a priori information about the geometry of the space. The algorithm is tested on INRIA human database where superior detection rates are observed over the previous approaches.

Detection and Segmentation of Multiple, Partially Occluded Objects by Grouping, Merging, Assigning Part Detection Responses

“…However, most recent methods use local features, because the local features are less sensitive to occlusions and other types of partially missing observations. Some examples are the wavelet descriptors in Schneiderman and Kanade (2000), the Haar like features in Viola and Jones (2001), the sparse rectangle features in Huang et al (2006Huang et al ( , 2007, the SIFT like orientation features in Mikolajczyk et al (2004), the Histogram of Oriented Gradients (HOG) descriptors in Dalal and Triggs (2005), the code-book of local appearance in Leibe et al (2004Leibe et al ( , 2005, the boundary fragments in Opelt et al (2006), the biologically-motivated sparse, localized features in Mutch and Lowe (2006), the shapelet features in Sabzmeydani and Mori (2007), the covariance descriptors in Tuzel et al (2007), the motion enhanced Haar features in Viola et al (2003), the Internal Motion Histograms (IMH) in Dalal et al (2006), and the edgelet features used in our previous work Nevatia 2005, 2007c). The above features are mostly shape oriented, because shape is the most consistent and salient image cue for many object classes.…”

“…Viola and Jones 2001;Huang et al 2007;Sabzmeydani and Mori 2007;Tuzel et al 2007;Dalal and Triggs 2005;Viola et al 2003;Zhu et al 2006;Schneiderman and Kanade 2000) learn object classifiers, whose input is a rectangular image subwindow and whose output is a prediction of the presence/absence of an object in this window. To capture the salient image characteristics of the objects class, a large variety of image features has been developed.…”

“…Most of the non-object sub-windows are rejected by the first few layers. The layer classifiers can be learned by different learning methods, such as an SVM in Dalal and Triggs (2005), Dalal et al (2006) and AdaBoost in Viola and Jones (2001), Viola et al (2003), Huang et al (2007), Zhu et al (2006), Tuzel et al (2007), Sabzmeydani and Mori (2007).…”

“…In recent years, methods for direct detection of objects have become popular, and promising results have been achieved for several object classes, including faces (Viola and Jones 2001;Huang et al 2007), pedestrians Nevatia 2005, 2007c;Sabzmeydani and Mori 2007;Tuzel et al 2007;Gavrila 2007;Viola et al 2003;Zhu et al 2006;Mikolajczyk et al 2004), and cars (Wu and Nevatia 2007a;Schneiderman and Kanade 2000). In these methods, no prior segmentation is applied; rather, the image is scanned by windows of various size and a determination as to the presence of the desired object is made in this window.…”

Fast Human Detection in Surveillance Video

“…Many supervised learning algorithms have proposed for detecting human object from an image [3,4,5,6]. Tuzel et al [7] used covariance matrices as object descriptors.…”