In this paper, for the reduction of the computation time of a deformable approach to pattern recognition, prototype-parallel displacement computation on GPUs (PPDC-GPU) is proposed. The displacement computation used in this study has the virtue of simplicity and consists of locally parallel processing, therefore it is expected to be suitable for the implementation on graphics processing units (GPUs). In the proposed method, large plates of image and displacement function are generated from input images, prototypes, and displacement functions on the main memory, and then these plates are transferred collectively to the video memory. After the parallel computation of displacement between the input image plate and the prototype image plate on the GPU, the displacement function plates are transferred back to the main memory. The simulation results show that PPDC-GPU reduces the computation time to less than 10% of the ordinary implementation on CPUs. This study especially focused on handwritten character recognition, since it is one of the most fundamental and important problems in the field of computer vision and pattern recognition. However the proposed framework can be widely applied to other problems, for instance, face recognition or object recognition.

View full abstract

This article addresses a method that predicts colorimetric values of an object under an arbitrary illuminant from an RGB image photographed by a digital still camera. When using a 3-band camera, multiple reflectance spectra that yield different colorimetric values may give the same sensor responses. This phenomenon implies that a conversion from sensor responses to accurate colorimetric values is a one-to-many relationship and cannot be uniquely determined. The proposed method can derive the colorimetric values with the highest degree of expectation from the given sensor responses by referring to the statistic feature of reflectance spectra of objects with the same material as the object to be photographed.

View full abstract

Part I of this article proposed a novel method that predicts colorimetric values of an object under an arbitrary illuminant from an RGB image photographed by a digital still camera. In this paper, the performance of the proposed method is evaluated. Computer simulations were carried out to evaluate the color prediction accuracy for various combinations of illuminants, camera sensitivities, object categories, and population data used for constructing color prediction models. The results show that the proposed method outperforms conventional methods in terms of both accuracy and robustness against selection of population data. An experiment using a digital camera image also showed that the proposed method is of practical use for actual images.

View full abstract

The Number of breathing problems (SDB: Sleep disordered Breathing) has increased during sleep in recent years. SDB causes serious diseases such as diabetes and high blood pressure, etc. Many kinds of methods to observe the breathing state are developing. However, devices that have already been commercialized and used actually in the hospitals are the contact types, which cause people to feel pain or discomfort. So, we developed a system which displays breathing movement with an indication of rough physical shape using Fiber Grating 3D sensor. This device can, in real time, monitor the body and the breathing activity. At a glance, we can recognize an abnormality in movement when there is abnormal breathing. Also we can measure quantity of volume change of a body from breathing movement quantitatively. Thus, we show that we can measure the breathing movement using this system.

View full abstract

This paper proposes Player's View as a new presentation method of Free Viewpoint Video, and a method to estimate trajectories of soccer players for generating Player's View. We estimate the trajectories of soccer players by using multiple cameras set in high place to reduce occlusion. Shadow and Player regions extracted from the images are used for player's position estimation. Furthermore, our method utilizes color information of the players' uniforms to improve the accuracy of object tracking. We have applied our proposed method to a real soccer game and demonstrated its effectiveness. Player's View was generated from the trajectories successfully.

View full abstract

We propose an automatic 3D human head modeling method using both a frontal facial image and geometry. In general, template mesh fitting methods are used to create a face model from a facial range data obtained by range scanner. However, previous fitting techniques need to manually specify markers to the scanned 3D geometry and to manually correct the 3D geometry of the missing parts that it is impossible to accurately measure the head geometry of the hair region. We therefore complement this region's 3D geometry with the template mesh's one. Our technique can generate the head model that the scanned 3D face geometry and the template mesh's one are seamlessly connected. The computational time of our method is much faster than previous template mesh fitting methods. We therefore conclude that proposed method is effective to create a large amount of head models in game and film industry and an entertainment system.

View full abstract

The aesthetic curves include the logarithmic (equiangular) spiral, clothoid, and involute curves. Although most of them are expressed only by an integral form of the tangent vector, it is possible to interactively generate and deform them and they are expected to be utilized for practical use of industrial and graphical design. However, the curvature of a log-aesthetic curve segment must be monotoinically increasing or decreasing and a strong constraint as a function of arc length is imposed on its curvature. For geometrical design, it is desirable not to impose strong constraints on the designer's activity, to let him/her design freely and to embed the properties of the log-aesthetic curves for complicated curves with both increasing and decreasing curvature. Hence we develop a discrete filter named the discrete log-aesthetic filter to fair a sequence of points with noises such as measurement data used in reverse engineering and to fit it locally to log-aestehtic curves. Furthermore we disucuss how to extend it for space curves and surfaces.

View full abstract

Rendering scattering materials is an important issue in realistic rendering. This paper describes calculation and interpolation methods of BRDFs/BTDFs (Bi-directional Reflectance/Transmittance Distribution Functions) of scattering materials. We introduced the plane-parallel scattering theory, which enables us analytic calculation of BRDFs/BTDFs of uniform scattering materials. We analysed calculated BRDFs/BTDFs and proposed their interpolation methods against scattering parameters and width of scattering materials.

View full abstract

We propose a method for avoiding the statistical bias of Metropolis light transport with multiple importance sampling based on the primary sample space. The statistical bias produces incorrect results which are much brighter or darker than true result in physically based rendering. This problem has been ignored because it appears only in special scenes. We solve the statistical bias problem by using two image buffers for storing sampling results.

View full abstract

This paper proposes a method for suppressing rapid volume-loss of nucleate boiling simulation. The method suppresses volume loss by use of the CIP method for a conservative system for advection of the liquid interface. Furthermore, we propose a technique for nucleate boiling simulation using the “nucleation solid,” defined as an object that is not registered to the simulation engine for collision detection and is not rendered. Here, we introduce a method to reduce volume loss induced by nucleation in existing boiling or bubble simulations. To solve the problem of volume loss in nucleate boiling simulation, we set up the inlet boundary condition at the wall surface, and control fluid flow by the nucleation solid object and implement nucleation as an operation of “pushing fluid out of the wall surface” by a solid object.

View full abstract

In movies or games, physically-based simulation has become an indispensable technique to generate realistic animations. In movie productions, the scene is designed beforehand in a storyboard, and while keeping a realistic appearance, there is often a need of non-realistic behavior. During the movie production, the control of the behavior of the objects is slightly possible by adjusting certain parameters of the simulation such as materials. Despite that, setting them to obtain the expected results is a laborious and not a straight-forward process. In this paper, we propose a method for the destruction simulation based on a mass-spring model, where objects are represented as set of elements interconnected by spring, to control the shape of the objects after destruction. We control the shape resulting from destruction by letting the user to specify the hard and the fragile parts of the objects. Based on this information, the toughness of the springs are automatically adjusted. As a result, we can generate realistic destruction animations and at the same time control the shape of the objects after destruction.

View full abstract

In order for volume rendering to convey the features of a given volume dataset effectively, it is required to design the illumination environment with which we can provide enriched depth cues for the rendered features, as well as to specify the transfer functions which emphasize the features, to locate the viewpoint in such a way as to minimize the occlusion artifacts among the projected features. In this paper, we strive to extend the conventional illumination entropy-based framework for surface models to come up with a novel method for finding an optimal location of light source for further enhancing differential topology features embedded within the volume. The feasibility of the proposed method is empirically proven through several experiments with practical datasets for the case with a single parallel light source. The extensibility of the method to the problem of multi light sources placement is also discussed.

View full abstract

We propose a general-purposed parallel algorithm for volume segmentation, which does not require any prior knowledge on volume nor region. The algorithm provides binary tree structured split-and-merge mechanism to search and localize boundaries along discontinuities and adapts the partition of volume to those detected discontinuities. This algorithm is independent from order of processing or seed selection and maintains consistencies of adjacent process blocks without their interprocess communication. Our efficient hierarchical step-wised mechanism in merging target evaluation makes the merge process so simple and efficient that only two blocks with the same parents are considered at each merge step in binary fashion.

View full abstract

Produced are varieties of information, such as tracking, speed, and recognition of cars as well as traffic density from trafiic video, while there are many tasks to be solved. As an example, it is important but not easy to separate a car and shadow attached to it since a moving area includes the car and the shadow when lit by the sun. This paper presents a novel method for car shadow detection with improvement for background update. In general, an intermediate part, called penumbra, exists between shadow and lit parts. Luminance gradually changes in the penumbra but little changes in the shadow area. This leads us to a new shadow model, i.e. a shadow part consists of a concavity on the left edge, a convexity on the right edge, and very little luminance change between the two edges after differentiation of the luminance value of the shadow part. This shadow model is enhanced with another property, i.e. width of a shadow coincides with that of a moving area in the shadow part. Based on the new shadow model, an experiment is carried out showing its superior performance in shadow detection rate, where the rate is 83%, much greater than 70∼78% by other methods proposed in the past.

View full abstract

Computer-generated hologram is constructed for three dimensional image reconstruction of a virtual object. Error diffusion method is used to improve the reconstructed image. The selection of diffusion coefficient exerts an influence to the image quality of the reconstructed image. Therefore, many researchers study the optimization of coefficients. In this paper, estimation of coefficients using statistical analysis is done. Using regression analysis, it's possible to relate coefficients and object position. Calculation time and reconstructed image quality of proposed method are better than GA and SPA.

View full abstract