Atmospheric phenomena such as rainbows and halos are caused by interference, diffraction and dispersion of light. To render the kinds of atmospheric phenomena, the spectrum of light should be taken into account because the optical phenomena greatly depend on the wavelength. Traditionally, rainbows are rendered by means of geometric optics. These methods, however, cannot render supernumerary rainbows and visual variations of rainbows depending on the radius of water drops. The paper proposes a method for rendering rainbows taking into account wave optics and spectral distribution of light based on physical laws. The method is taking into account the distribution of water drops, such as the radius and the density of water drops in the atmosphere. And also we are taking into account atmospheric light scattering by air molecules and aerosol. These efforts enable our method to render the visual variations of rainbows and supernumerary rainbows. It helps visual understanding of rainbows and atmospheric conditions in meteorology.

View full abstract

In this paper, we propose a particle-based method for real-time flame rendering. In the field of visualization, particle-based volume rendering has been introduced for the purpose of previewing. Our method is a reasonably-simplified version of the particle-based volume rendering. Compared with previous flame rendering methods, both detail representation and controllability are sufficiently improved. In addition to this, it is easily implemented using standard APIs, and its fast enough rendering speed allows us to adjust environmental settings on the fly. Moreover, our method requires a few standard parameters as particle attributes, and thus it would be used in conjunction with existing particle simulators.

View full abstract

Surface deformations based on physically-based simulations are used to represent elastic motions such as human skins or clothes in the field of 3DCG applications. SM (Shape Matching)^2),6) has particularly attracted attention as a fast and robust method which achieves elastic-like motions. However, the original LSM deformation method generates far from realistic motions especially when stretching an object, because of not preserving its volume. We propose a novel volume-preserving deformation method based on LSM. Our method can achieve more elastic-like motions than the original LSM while keeping fast and robust computations. And, We show an adaptableness for arvitrarily distributed points, as well as latticed points.

View full abstract

In this paper, we propose a robust projection operator for noisy point clouds with outliers. The operator is an extension of Moving Least Square (MLS) projection and inherits its smoothing property. A key idea is to compute average positions and projection directions to prevent the influence of outliers. This is done by a kind of median-filtering operation. In all our experiments, our operator has been able to eliminate outliers well compared with previously-proposed projection operators, and is also able to reconstruct clean and smooth meshes from filtered point clouds.

View full abstract

We propose a method for rendering realistic three-dimensional raining scene. The proposed method gives two main contributions to rendering raining scene. The first one is to propose a simple model that deals with temporal-spatial localities such as wind effect, density and intensity of rainfall, and raindrop brightness, so as to represent environmental conditions that differs on a scene-to-scene basis. The other one is to propose a raindrop trajectory computing method whose computation load immunes to the number of raindrop, wind effect and complicated camera movement. The usefullness of the proposed method is demonstrated through some examples.

View full abstract

This paper describes an interactive technique for generating brush strokes used in exemplar images. The proposed method decomposes several exemplar images of brush strokes into small fragments that are stitched together through one-dimensional texture synthesis in order to generate the final brush stroke silhouettes. In contrast to previous methods in which the strokes are generated by per-pixel computation or using physically based simulation, our method is not only less computationally expensive but also capable of simulating various kinds of brushes. In addition, we introduce a fluid-dynamics-driven post-process and a Kubelka-Munk synthesis method to make the paint strokes appear more real. Our evaluation experiments show that our method can handle various media such as oil paints, watercolors, charcoal, and pastels. We also obtained interesting results with lipsticks and fingers.

View full abstract

Pedestrian protection is a crucial component of driver assistance systems. Considering all the available cues for predicting the possibility of collision is very important. The direction in which the pedestrian is facing is one of the most important cues predicting where the pedestrian may move in future. In this paper, we address the problem of single-frame pedestrian orientation estimation in real-world scenes with a single camera from a moving vehicle. We present a novel approach to determine the pedestrian' orientations using cascade orientation estimation, by exploiting the head orientation by a multi-Bayesian model.
Experiments on a large amount of real-world data show a significant performance improvement in octant orientation estimation of about 64% accuracy by head orientation consideration, compared to one vs. one and one vs. all multi-class classification methods.

View full abstract

We propose a vision-based Human-Computer Interface for Text Entry. The system uses a web camera for detecting tongue protrusion gestures, which are interpreted as the signals of International Morse Code. These gestures can be generated independently; hence, the traditional 3:1 ratio between dashes and dots can be disregarded. Employing Morse code for text entry requires memorizing it in advance. In this paper, users are provided with a Visual Chart where input characters are displayed on the screen. Navigating the chart in order to select a character matches perceptually the position of the tongue gestures. Thus, without previous knowledge of Morse code, users are able to start typing just by looking at the screen. Furthermore, the proposed interface allows learning the code by associating it with the tongue actions, while already obtaining tangible results. The Text Entry protocol consists of Timers that can be adjusted according to the level of expertise of users. The best text entry rate obtained was 2.54 WPM. We also provide a method to calculate theoretical speeds, which indicate the lower bound for the speeds obtained in practice. Finally, the Visual Chart contains 30 characters; however, it is possible to expand it in order to encode more information while maintaining the same text entry protocol.

View full abstract

We propose a registration method for cross-sectional images to reconstruct a three-dimensional shape of an ocular fundus. Cross-sectional images obtained by an Optical Coherence Tomography are displaced because of an eye movement. The proposed method registers cross-sectional images adjusting parameters of Roll-Pitch-Yaw rotations based on the eye movement. In addition, we apply an objective function considering the shape of an optic disc to reconstruct an ocular fundus accurately. The proposed method reconstructed an accurate shape of the ocular fundus comparing to a conventional method.

View full abstract

In this paper, we propose a graph-based calculation method for temporally stable superpixels. A superpixel, that is a small region with similar colors and textures, is recently used to improve the recognition accuracy and reduce the processing time in object recognition. When superpixels are calculated from the video, the results should have temporal stability. To achieve the above property, we develop a new calculation method based on the conventional method. Our method can calculate the temporally stable superpixels by estimating the corresponding points. In quantitative experiment using the artificial images, proposed method is more stable than the conventional methods.

View full abstract

Template matching is an effective method for calculating the similarity between corresponding-points. However, a mismatching-problem is caused in this method due to the occlusion, shadow, or distortion of objects. In the conventional method for calculating the similarity between corresponding points, a fixed template size is applied for template matching. Therefore, the information in the fixed templates may be excess or insufficient. In fact, if the template size can be adapted conditionally on the calculation of retrievals, we can realize accurate systems. This paper shows that there is an optimal template size for every searched image based on the residual probability distribution function and proposes an automatic adaptation method for the template size. The proposed method depends on the fact that the distribution of the residuals can be fit to the χ² distribution and applied to the selection of the residual between the reference image and the input image.

View full abstract

In this paper we present a novel method for the registration of point cloud data obtained using terrestrial laser scanner (TLS). The final goal of our investigation is the automated reconstruction of CAD drawings and modeling of objects surveyed by TLS. Individual TLS point cloud needs to be registered to the same coordinate system, because the object is scanned from multiple positions. We have developed an automated feature extraction method. The proposed method can calculate the rigid parameter effectively by utilizing the position and normal vector of the extracted plane instead of dealing with plenty point cloud directly. The result of some fundamental experiments shows the effectiveness of the proposed method.

View full abstract

In this paper, we propose an improved feature selection method based on genetic algorithms (GA) that selects not only features but also training samples when creating reference patterns. In this method, the chromosome is composed of a set of bits that represents the state of feature selection. The reference pattern in each category is created from the selected training samples, and the discrimination rate for all training samples is used as the fitness value in GA. GA is used as a tool to find the suboptimal combination that increases the discrimination rate from the enormous number of combinations of features and training samples. We perform experiments using five sets of hand-written Kanji character patterns. Each set consists of patterns in two similar categories. The results show that the proposed method, which selects features and training samples, provides higher recognition rates than the conventional method that selects only features.

View full abstract

Personal identification by handwriting in forensic science (scientific criminal investigation) is a typical application field of the text-dependent and off-line writer verification. However, it still stays at the research and development stage, and it has not arrived at the situation of being practicably used in the handwriting identification. This paper proposes a simple writer verification method using the conventional maximum likelihood classifier and based on layout pattern of feature points which are composed of start, corner and end point of strokes. These feature points are fundamental points which represent shape of Chinese characters, and are stably extractable. Though only a limited feature for handwriting is utilized in the proposed method, verification experiments on the samples of 320 writers showed that verification rate was improved to practical use level by introducing uniform geometric transform for all feature points as a preprocessing in order to restrain intra-individual (within writer) variation.

View full abstract

Using color information can significantly improve the face recognition rate instead of using the grayscale luminance image. However, there are few works that try to compare the color space models on face recognition. In this paper, we investigate thirty different color space models on face recognition using the classical principal component analysis (PCA). Through the extensive experiments we find that after successfully diminishing the influence of the illumination the recognition accuracy can be improved by 4.6∼5.5 percent points.

View full abstract

When the contrast of an image provided by photography is not good, a method to improve contrast by the software is effective for the image. The linear gray-level conversion and the histogram equalization are proposed as the conventional contrast improvement methods based on the distribution of gray-level included in the input image. However, because the conventional method does not use a spatial feature included in an image, contrast improvement is insufficient and unnatural. There is the method for contrast improvement of the image based on the differential gray-level histogram. This method can get a good result for the whole image, but it is sometimes arisen that the details of the image deteriorate. The objective of this research improves contrast of both the whole image and the image details. This paper proposes a spatial characteristic, the image segmentation, linear interpolation in order to improve contrast method. First, I divide an image into some blocks. I generate the Look-Up-Table from the differential gray-level histogram in each block. Finally, the output gray-level of each pixel is determined based on linear interpolation from the vicinity of the four LUT. The experimental results show that the image had good and natural contrast by the proposed method.

View full abstract

As for acquisition and presentation of the remote place sight by telepresence, it is necessary to show the sight by picture which is expected to look freely and stereoscopic vision of freedom. As the conventional technique, there were method which use two or more cameras to take pictures, and another method which use several pictures to generate the middle viewpoint. However, the former, it tends to become an expensive and complicated system, and the latter, there is a problem that information is missing with occlusion. In this paper, we propose a system which used one camera and a spin mirror to generate the free viewpoint 3D image. We checked the validity of proposal system by the experiment which used the plot type system.

View full abstract

We describe an image haptization method based on color and edge information. Setting an impedance information is necessary to touch an image by using a haptic device. However, there is no method for automatically setting the parameter values to achieve the natural feeling by using only the image information. Thus, we propose a method for generating parameter values by using the color information of the image. In addition, we propose a method that can improve recognizing an image by using the edge information of the image. Finally, we developed the image haptization system, and verified the effectiveness of the proposal method.

View full abstract