K. Yow and R. Cipolla. Finding initial estimates of human face location. In In Proc. 2nd Asian Conf. on Comp. Vision, volume 3, pages 514--518, Singapore, 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....Today, the problem is still not solved in its most general case and many competing techniques are proposed. In this work, we show how to combine different algorithms using a Bayesian Network to efficiently propagate evidence gained by various techniques that have been approved in the past [1, 2, 5, 6]. 1 CONTENTS 2 Contents 1 Introduction 4 1.1 Definition of Task to solve . 4 1.2 Questions to be answered in this Report . 4 1.3 Relevance of Face Detection . 4 1.4 Vocabulary used . ....

Kin Choong Yow, Roberto Cipolla: "Finding Initial Estimates of Human Face Location", Proceedings 2nd Asian Conference on Computer Vision, vol. 3, pp. 514-518, Singapore, 1995.

....Also, due to occlusion or missing features (eyebrows, usually) we need to decompose the face model into components consisting of 4 features, which are common occurrences of faces under different viewpoints or different identity. These groups are called Partial Face Groups or PFGs (Yow and Cipolla [33]) These PFGs are further subdivided into components consisting of 2 features (horizontal and vertical pairs Hpair and Vpair) fig. 1) for the purpose of perceptual grouping and evidence propagation. In order for feature detection to be robust we have to use image features that are invariant ....

....in the belief network is the prior probability of the root node (the face node in this case) and this is often hard to estimate. Certainly, the choice of an appropriate prior depends on the complete space of hypothesis. We may assume an uniform prior for our case. In our previous approaches ([33], 34] we used a belief network comprising of 4 child nodes, one for each of the 4 partial face groups (fig. 5(a) This was shown to be highly effective for fronto parallel view of faces because all 4 PFGs can be detected in this view, giving a large amount of evidence for true face candidates. ....

[Article contains additional citation context not shown here]

K. C. Yow and R. Cipolla. Finding initial estimates of human face location. In Proc. 2nd Asian Conf. on Comp. Vision, volume 3, pages 514--518, Singapore, 1995.

....in terms of conditional probabilities (Sarkar and Boyer [10] Each node has a conditional probability table (CPT) associated with it, describing the conditional probability of each value of the variable, given each possible combination of the values of the parent nodes. In Yow and Cipolla [14], the use of belief networks was proposed for defining the probability for each face candidate. The belief network used is shown in fig. 5(a) A uniform prior is assumed, and the CPT entries estimated directly using the statistics of the set of examples (Russell et.al. 9] TopPFG BottomPFG ....

K. C. Yow and R. Cipolla. Finding initial estimates of human face location. In Proc. 2nd Asian Conf. on Comp. Vision, volume 3, pages 514--518, Singapore, 1995.

....view) and thus it is difficult to extend this approach to multiple views. The feature based approach searches the image for a set of facial features and groups them into face candidates based on their geometrical relationship. Leung et.al. 7] Sumi and Ohta [12] and Yow and Cipolla [17] reported work using this approach. Though this approach can be easily extended to multiple views, it is unable to work well under different imaging conditions because the image structure of the facial features vary too much to be robustly detected. The approach based on neural networks detects ....

....image. 5.1. Preattentive feature selection The preattentive feature selection stage is performed in two steps. First, a list of interest points is found from the image. This is achieved by filtering the image using a preattentive filter (a second derivative of Gaussian used in Yow and Cipolla [17]) and then searching for local maxima. Next, the edges around each interest point are examined. Similar edges are linked using a boundary following algorithm. If we find the existence of two roughly parallel edge segments with opposite polarity on both sides of the interest point, then this ....

[Article contains additional citation context not shown here]

K. C. Yow and R. Cipolla. Finding initial estimates of human face location. In Proc. 2nd Asian Conf. on Comp. Vision, volume 3, pages 514--518, Singapore, 1995.

....of the right scale and orientation (efficiently implemented by using steerable scalable basis filters Perona [11] Freeman and Adelson [4] The detected facial features are then grouped into face candidates according to their geometrical relationships (Leung et.al. 8] Yow and Cipolla [17]) The difficulty, however, in using this approach is that the feature extraction stage is either not robust enough or it detects too many candidate features. Chen et.al. 2] cope with viewpoint variations by performing matching in 3 views (1 fronto parallel and 2 profile views) using a fuzzy ....

....and reasoning activities on these points and regions based on the Gestalt laws (Kohler [6] Koffka [5] and on the model of the face. The first stage, the preattentive feature selection stage, finds a list of interest points by filtering the image with a matched bandpass filter (Yow and Cipolla [17]) and then searching for local maxima. Next, the edges around each interest point are linked using a standard boundary following algorithm. A feature point is found if there are two roughly parallel edge segments with opposite polarity on both sides of the point. The extent of the feature region ....

[Article contains additional citation context not shown here]

K. C. Yow and R. Cipolla. Finding initial estimates of human face location. In Proc. 2nd Asian Conf. on Comp. Vision, volume 3, pages 514--518, Singapore, 1995.

....Also, due to occlusion or missing features (eyebrows, usually) we need to decompose the face model into components consisting of 4 features, which are common occurrences of faces under different viewpoints or different identity. These groups are called Partial Face Groups or PFGs (Yow and Cipolla [34]) These PFGs are further subdivided into components consisting of 2 features (horizontal and vertical pairs Hpair and Vpair) fig. 1) for the purpose of perceptual grouping and evidence propagation. In order for feature detection to be robust we have to use image features that are invariant to ....

....in the belief network is the prior probability of the root node (the face node in this case) and this is often hard to estimate. Certainly, the choice of an appropriate prior depends on the complete space of hypothesis. We may assume an uniform prior for our case. In our previous approaches [34, 35], we used a belief network composed of 4 child nodes, one for each of the 4 partial face groups (fig. 5(a) This was shown to be highly effective for fronto parallel view of faces because all 4 PFGs can be detected in this view, giving a large amount of evidence for true face candidates. However, ....

[Article contains additional citation context not shown here]

K. C. Yow and R. Cipolla. Finding initial estimates of human face location. In Proc. 2nd Asian Conf. on Comp. Vision, volume 3, pages 514--518, Singapore, 1995.

No context found.

K. Yow and R. Cipolla. Finding initial estimates of human face location. In In Proc. 2nd Asian Conf. on Comp. Vision, volume 3, pages 514--518, Singapore, 1995.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC