Shun Ido

Face detection evaluation: a new approach based on the golden ratio {\Phi}

Face detection is a fundamental research area in computer vision field. Most of the face-related applications such as face recognition and face tracking assume that the face region is perfectly detected. To adopt a certain face detection algorithm in these applications, evaluation of its performance is needed. Unfortunately, it is difficult to evaluate the performance of face detection algorithms due to the lack of universal criteria in the literature. In this paper, we propose a new evaluation measure for face detection algorithms by exploiting a biological property called Golden Ratio of the perfect human face. The new evaluation measure is more realistic and accurate compared to the existing one. Using the proposed measure, five haar-cascade classifiers provided by Intel©OpenCV have been quantitatively evaluated on three common databases to show their robustness and weakness as these classifiers have never been compared among each other on same databases under a specific evaluation measure. A thoughtful comparison between the best haar-classifier and two other face detection algorithms is presented. Moreover, we introduce a new challenging dataset, where the subjects wear the headscarf. The new dataset is used as a testbed for evaluating the current state of face detection algorithms under the headscarf occlusion.

Independent components analysis-based nose detection method

Automatic detection of facial features plays an important role in many face-related applications. Among these features, nose region is the least varying part of the human face. In this paper, a method for nose region detection is presented. The method adopt Independent Components Analysis (ICA) as a subspace classifier to classify the face candidate region to nose or non nose. The ICA basis vectors are estimated by the FastICA algorithm. The training has been done using features of nose appearance and shape characterized by the edge information. The effect of preprocessing step on the performance at different dimensions of ICA subspace is also examined. The feasibility of the proposed method has been successfully tested using different databases under various imaging conditions and the results are encouraging.

Haptic simulation of rheological objects with Verlet integration

In virtual reality, educational simulators are expected to be enhanced the learning effect by representing haptic feedback, the feeling of touching objects, in the virtual world. The haptic feedback can represent more realistic feel of a material, such as the softness. For example, it is the feel of cutting organs in medicine and mixing foods in cooking. Therefore, the haptic feedback can produce experiences closer to reality, and more realistic learning effect can be expected in virtual world. However, deformable objects in the virtual world, such as organs and foods, are fractured by adding strong force generated by touching them, and normal simulations are difficult in the situation. Then, for solving this problem, we applied Verlet integration, whose example of the application is few in conventional haptic simulations, to calculations of deformable object positions for achieving break-proof deformable objects. For showing that deformable objects which are applied Verlet integration are break-proof, we compared the robustness of Verlet integration with other methods in the same situation where each method is applied to a deformable object. As the result, we verified the situation in which the only deformable object in which Verlet integration is applied is not fractured, and we showed that Verlet integration is more suitable for deformable object simulations than other methods are.

A robust method for nose detection under various conditions

In this paper, a robust fully automatic method for nose field detection under different imaging conditions is presented. It depends on the local appearance and shape of nose region characterized by edge information. Independent Components Analysis (ICA) is used to learn the appearance of nose. We show experimentally that using edge information for characterizing appearance and shape outperforms using intensity information. The influence of preprocessing step on the performance of the method is also examined. A subregion-based framework depending on statistical analysis of intensity information in the nose region is proposed to improve the efficiency of ICA. Experimental results show that the proposed method can accurately detect nose with an average detection rate of 95.5% on 6778 images from six different databases without prior detection for other facial features, outperforming existing methods.