Vasanth Philomin

An investigation into the use of partial-faces for face recognition

Even though numerous techniques for face recognition have been explored over the years, most research has primarily focused on identification from full frontal/profile facial images. This paper conducts a systemic study to assess the performance when using partial faces for identification. Our specific approach considers an ensemble of radial basis function (RBF) networks. A specific advantage of using an ensemble is its ability to cope with the inherent variability in the image formation and the data acquisition process. Our database consists of imagery corresponding to 150 unique subjects, totalling 3,000 facial images with /spl plusmn/5/spl deg/ rotation. Based on our experimental results, we observe that the average cross-validation performance is the same, even if only half the face image is used instead of the full-face image. Specifically, we obtain 96% when partial faces are used and 97% when full faces are used.

Visual Surveillance in Retail Stores and in the Home

This paper presents an overview of video surveillance activities at Philips Research USA, which serves as a research arm of Philips Communications, Security and Imaging, a global leader in the design and manufacture of advanced automatic security systems. We concentrate on two application domains: professional security market, with emphasis on retail monitoring, and a low-cost, automated residential security. The main application for pan-tilt-zoom (PTZ) cameras is typically tracking of intruders throughout the facility. In addition, camera calibration is an important capability, because it allows geometric reasoning with respect to the floor plan as well as enhanced control of the camera. Another major component of our work concerns video content analysis, the detection of security related objects and events from video. The system typically processes video in real-time and can provide immediate alarms to alert the security operator. Relevant information is also stored in a database so that it can be efficiently retrieved later. The third and final topic discussed in the paper is a residential monitoring application, namely an intruder detection system. We describe detection of moving objects robust to changes in lighting and an object classification scheme based on radial-basis networks.

Toward fully automatic object detection and segmentation

An automatic procedure for detecting and segmenting anatomical objects in 3-D images is necessary for achieving a high level of automation in many medical applications. Since todays segmentation techniques typically rely on user input for initialization, they do not allow for a fully automatic workflow. In this work, the generalized Hough transform is used for detecting anatomical objects with well defined shape in 3-D medical images. This well-known technique has frequently been used for object detection in 2-D images and is known to be robust and reliable. However, its computational and memory requirements are generally huge, especially in case of considering 3-D images and various free transformation parameters. Our approach limits the complexity of the generalized Hough transform to a reasonable amount by (1) using object prior knowledge during the preprocessing in order to suppress unlikely regions in the image, (2) restricting the flexibility of the applied transformation to only scaling and translation, and (3) using a simple shape model which does not cover any inter-individual shape variability. Despite these limitations, the approach is demonstrated to allow for a coarse 3-D delineation of the femur, vertebra and heart in a number of experiments. Additionally it is shown that the quality of the object localization is in nearly all cases sufficient to initialize a successful segmentation using shape constrained deformable models.

Disparity Estimation in Stereo Sequences using Scene Flow

Fang.F.Liu@philips.comPhilips Germany Vasanth Philomin Vasanth.Philomin@philips.comPhilips Germany Abstract This paper presents a method for estimating disparity images from a stereo image sequence. While many existing stereo algorithms work well on a single pair of stereo images, it is not sufficient to simply apply them to temporal frames independently without considering the temporal consistency between adjacent frames. Our method integrates the state-of-the-art stereo algorithm with the scene flow concept in order to capture the temporal correspondences. It computes the dense disparity images and scene flow in a practical and unified process: the disparity is initialized by a hybrid stereo approach which employs the over-segmentation based stereo and pixelwise iterative stereo; then the scene flow, estimated via a variational approach, is used to predict the disparity image and to compute its confidence map for the next frame. The prediction is modeled as a prior probability distribution and is built into an energy function defined for stereo matching on the next frame. The disparity can be estimated by minimizing this energy function. Experimental results show that the algorithm is able to estimate the disparity images in an accurate and temporally consistent fashion.

Illumination Invariant Face Recognition

This paper introduces a system in which the face recognition is carried out with the help of the local feature descriptors. As today face recognition is a very active area of research, a system which carries out the robust face recognition is needed to be developed. Face recognition is used in various fields of security and access control, surveillance military etc. here the system is developed in which face recognition is done under varying illumination (lighting) conditions. Local directional number pattern (LDN) is the descriptor which is used here to extract the features from the face image. The LDN produces the compact code which extracts the useful features of the face. The face image is divided into several regions and then for each region LDN is computed and then the histogram of each region is taken for the sake of comparison. Then all the histograms are concatenated to form a feature vector. The LDN thus perform the best under the varying illumination conditions for the correct face recognition. KeywordsFace recognition, Gabor filter, LBP (Local Binary Pattern), feature extraction, FRM (Face recognition method).