Motion representation is one of the most important elements in producing realistic virtual humans in computer graphics. Although manual production techniques, such as keyframe interpolation, enable animators to freely synthesize motion, manually creating motions as detailed as those obtained by motion capture requires substantial skill and time. Therefore, this paper proposes a method to obtain high-detail motion with high expressive power from inexpressive and low-detail motion. The method entails deep-learning based detailing of motion representations, by using a motion enhancement network built on a recurrent neural network with a long short-term memory. This method may potentially produce high-detail motion resembling actual human motions from arbitrary low-detail motions created by novice users using keyframing techniques. Our experiments demonstrate the synthesis of high-detail motion, including the amount of movement, from low-detail motion of walking and talking.

We propose a new corner estimation technique for 3D point cloud on convex polyhedral surfaces acquired by a 3D scanner. A corner is an important property of objects, and it is used in various technologies such as 2D or 3D object recognition system. The Harris corner detector and SUSAN are widely used as a method of finding corners in a 3D point cloud. In these methods, however, high processing cost will often be a problem. In this paper, we propose a new corner estimation technique using Delaunay tetrahedralization. This technique is able to estimate corners of convex polyhedral surfaces efficiently by using characteristic tetrahedrons generated by Delaunay tetrahedralization. In an experiment, we compared the proposed technique with the Harris corner detector. As a result of the experiment, for convex polyhedral surfaces used in the experiment, it is confirmed that the proposed technique can estimate corners with less cost than the half of the Harris corner detector without serious loss of the accuracy of each corner location.

Template matching is a technique for finding a part of reference image which matches a template image. This paper presents a new fast template matching algorithm which can detect the most similar position. In the proposed method, first, an effective initial threshold is calculated using Winner Update Algorithm. Next, very fast template matching is achieved by using this initial threshold in Multilevel Successive Elimination Algorithm. Furthermore, Brick Partitioning which is a new partitioning method is used to reduce the computational cost of comparing a template with each position within reference image. Experimental results indicate that the proposed method can search faster than previous methods.

Wavelength Division Multiplexing (WDM) is a technology which multiplexes optical carrier signals on a single optical fiber by using different wavelengths. Lightwave networks based on WDM are promising ones for high-speed communication. If network nodes are equipped with tunable transmitters and receivers, a logical topology can be changed by reassigning wavelengths to tunable transceivers of nodes. Network performance is influenced by the logical node placements. Therefore, an efficient algorithm to obtain the optimal node placement to achieve the best network performance is necessary. In this paper, an iterated greedy algorithm is proposed for this node placement problem. The proposed iterated greedy algorithm consists of two phases, construction and destruction phases. As a local search algorithm, variable depth search is applied after the construction phase. The computational results showed that this iterated greedy algorithm outperformed the best metaheuristic algorithm for this problem.

In this paper, we proposed a method for solving the rectangle piece jigsaw puzzle assembly problem. A shape of a piece is a rectangle, and a puzzle image is RGB full color. The assembly of the puzzle is performed only using information of the pixel value on the border line of the piece. Pieces are connected by a matching function between two pieces. Not only the best matched piece is connected. The matching values to other pieces are used in the evaluation between two pieces. A simple method of puzzle assembly is that a single piece is connected to a block which is defined as a group of connected pieces. But types of piece combinations are restricted in the simple method. In the proposed method, each block is connected each other. Therefore, the best matched connection is selected from many types of combinations between pieces and blocks. The proposed method correctly assembled all pieces in 16 × 12-piece puzzles.

The human visual system can interpret two-dimensional (2-D) line drawings like the Necker cube as three-dimensional (3-D) wire frames. On this human ability Thomas Marill presented two important papers. First one proposed the 3-D interpretation model based on the principle to minimize the standard deviation of the angles between line segments in 3-D wire frame (MSDA), and reported the results of simulation experiments. Second one proposed the principle to minimize the description length on the internal representation in visual system. Motivated by Marill's principle to minimize the description length, we propose a principle to minimize the entropy of angle distribution between line segments in a 3-D wire frame (MEAD), which is more general than the MSDA one. And we implement the principle MEAD using a genetic algorithm (GA) as a simulation program. The results of simulation experiments show that the proposed principle of MEAD is more appropriate than the MSDA and another principle.

This paper proposes a model-based method for estimating the pose of a 2D articulated object from a single silhouette image using a genetic algorithm (GA). In this study, a human body viewed sideways is treated as a 2D articulated object. The model of the articulated object is given as a stick one. Its pose is represented in 2D position of the main stick and angles of all sticks connected at joints. The parameter space for representing the pose of the stick model is very large. Therefore, we limit the search space to the range in which the stick model does not protrude from the silhouette. For the search space limitation, the region of position and angle of the stick perfectly included in the silhouette is obtained for each stick in advance. The pose of the stick model is evaluated in the sum of the proximity measure from each pixel in the silhouette to the stick model. Experiments with synthetic silhouettes by a human figure design tool show that the proposed method can estimate various poses and that it can be applied to several human figures with a single stick model.

In this paper, we propose a method of the model-based pose estimation in the 3D world from a 2D image. This process consists in searching for the best value of coordinates (x, y, z) and of rotating angles (/spl theta//sub x/, /spl theta//sub y/, /spl theta//sub z/) in which the model object matches most exactly the given input edge image. Taking the match of objects as an index on the six-parameter space, this process can be regarded as a maximum searching problem. We use a genetic algorithm in the estimation of those parameters, and propose a new concept of fitness which takes edge direction into consideration. Our experiments show the results of the proposed method on edge input images.