ABSTRACT. This paper presents a method for measuring working hours of employees who work from home using a PC. In monitoring them by webcam, since their faces often look small in online meetings, it is difficult to apply regular face recognitions. The proposed method captures a frame per min. to a user and produces a black-and-white image recognizing the skin. Then, the image is discriminated into working in progress or not by utilizing run-lengths. The time is measured by counting the continuous frames discriminated as working in progress. Results conducted to 12 participants showed the discriminant ratios were 85% or higher.

ABSTRACT. This paper is proposed the fingering discrimination system that automatically discriminates whether the fingering is input according to the correct home position, and the typing learning game that can correct typing for the subjects. Using the fingering data of the subjects, it becomes possible to learn efficiently. The 15 subjects inputting 1800 characters in this experiment. The result showed that the discrimination accuracy of the proposed system was 98.8% and the typing learning game improved fingering by 25.6%. It is clear from the result that the discrimination accuracy was sufficient to evaluate the fingering of the subjects.

ABSTRACT. Utilizing deep learning in sports match analysis is very promising for efficiency, and there have been many studies focusing on volleyball. However, previous studies have used only the frames where the ball is alive (\ie, in-play frames), and recent data analysis methodologies rely on visual data, which are highly dependent on the position and angle of the camera relative to the court. As a result, these methods require a large dataset of images taken from various angles to improve accuracy. To efficiently extract and analyze the play as it is happening, we propose a model that distinguishes between in-play and out-of-play by combining visual data and audio data of volleyball match videos using late fusion. To investigate the effectiveness of the proposed model, we perform Grad-CAM visualization to determine which pixels the proposed model is focusing on.

ABSTRACT. This paper proposes a method that can interpret an evaluation criteria necessary for evaluating sport motions for achieving an automatic evaluation system using machine learning. For this, we build a running foot strike classifier using accelerometers as input and detect the contribution points for classification using the contribution degree of input values (CDIV). Next, we compare the CDIV with a kinematic model concerning ground strike motions. Then, we interpret the evaluation criteria of the classifier, which evaluates the acceleration peak that appears around the time at which the vamp of the foot leaves from the ground. As a result, we confirm other criteria is related to swinging the leg, which cannot be explained by the kinematic model.

ABSTRACT. This study aims to effectively design an apparel item by converting a polygon model of the human body obtained through 3D scan into a volume model. From the model, we can quickly generate a curved surface that retains a constant distance from the body surface, and thus design clothing material that ensures accurate “allowance” by defining the distance field for the body surface. The curved surface as an ideal garment that envelops the body can be generated by tweaking the threshold value for an isosurface from the body surface. Furthermore, by drawing a design line on the curved surface, the pattern required for actual sewing can be designed in three dimensions. The curved surface trimmed by the design line is converted into a polygon mesh. A dynamics simulation is applied to smooth the curvatures so that the needed flat surface pattern can finally be obtained. As a result of actually sewing the virtually designed pattern and try-on, it was proven that we could generate clothes with the appropriate allowance and proportions.

ABSTRACT. In recent years, the number of polygons that can be displayed in entertainment contents such as movies, full 3DCG animations, and games has been increasing, and a high level of detail is required for many background models. As a result, securing the volume and quality of the required CG models and managing the vastly expanding production pipeline have become issues. In particular, when the setting is a city, a large number of very detailed building models and other models that make up the cityscape are required. The creation of each of these models takes a great deal of time and effort, and has a significant impact on production costs. To solve this problem, procedural modeling, which can create models semi-automatically based on rules for the generation of models, has been introduced. The advantage of procedural modeling is that by creating a model under a set of rules, it is possible to create a model with a different shape by gradually changing the rules when the model is duplicated. In addition, when re-creating a model, it is possible to adjust the shape of the model to obtain the desired result by simply changing the numerical values, without having to start from scratch, thus saving time. The procedural modeling method of Japanese castle cities proposed in this paper enables the creation of models without knowledge of castles, and can be used for various contents in the video and game industries.

ABSTRACT. In this paper, we present an interactive modeling system for Japanese castles. We develop an user interface that can generate the fundamental structure of the castle tower consisting of stone walls, turrets, and roofs. By clicking on the screen displaying the 3D space with the mouse, relevant parameters are calculated internally, and the user can create 3D models of the Japanese-style castles. We apply some unique curves that often appear in ancient Japanese architecture. Our system allows the user to intuitively generate the castle models that are similar to those observed in the real world.

ABSTRACT. To create realistic images of outdoor scenes using computer graphics techniques, synthetic clouds with high-resolution details are required. However, the computation cost for generating high-resolution clouds is very expensive and time-consuming. In this paper, we address this problem and propose an efficient method to create high-resolution images of clouds. The method combines low resolution 3D simulation with high resolution 2D simulation. The low resolution 3D simulation is fast and is used for determining an overall shape of the clouds. Then the high resolution 2D simulation uses the result of the 3D simulation and adds small-scale details to them. We demonstrate that our method achieves ten times faster computation to generate realistic high-resolution images of clouds.

ABSTRACT. One of the significant characteristics in 3D point clouds that affect conventional registration algorithms and descriptor-based methods' performance is the overlapping ratio. As it decreases, it makes the resulting alignment worse - becoming a challenging task. In this work, we propose a pairwise registration method that, through a multi-scale and piece-wise search pipeline, achieves alignments close to the optimal without estimating key points and point descriptors in 3D point clouds with overlapping ratios below 40%. We comprehensively show our approach's results with different values of its parameters and core mechanism on reconstructed model data.

ABSTRACT. Registration is a fundamental technique in 3D point cloud processing which consists in correctly overlapping and integrating multiple 3D point clouds. Typical methods are ICP and feature point matching, but the former is computationally expensive and the latter strongly depends on the performance of feature points. On the other hand, VKOP, which finds feature points at virtual locations, has been found to have high performance in environments with many planar surfaces, but no registration method with it has been proposed yet. In this paper, we propose a new registration method that utilizes the optimal LRF of VKOP. From experiments, we confirmed that the proposed registration method can register more than twice as accurately as conventional methods with the same processing time.

ABSTRACT. In this study，we propose a method for reconstructing a 3D model from a moving image sequence using optical flow．In this method，the amount of movement between frames is first calculated using optical flow，and the overlap rate between frames is calculated from the obtained amount of movement．Next，the 3D model is reconstructed by selecting only the frames when this overlap rate falls below a certain value．We verify the effectiveness of the proposed method by experiments using some moving image data．

ABSTRACT. We propose an efficient semantic segmentation method for a large-scale point cloud. It is difficult to directly apply previous point-based semantic segmentation methods to large-scale point clouds because of their redundant structures. To solve this problem, we propose a novel point-based approach that predicts class labels for a downsampled point cloud and expands the labels to the whole point cloud by nearest-neighbor interpolation. The key idea of our approach is to give local features derived from whole point cloud to each sample point and convert those features into wider context features by a point-based model for small-scale point clouds.

ABSTRACT. In this paper, we present a method for hair plastic deformation caused by interaction with liquid. Because of its complexity, a hair shows characteristic behavior if it interacts with other materials, such as water. We focus on shape deformation of the hair caused by interaction with liquids considering an internal bond inside the hair. We introduce a method to switch between plastic and elastic deformation depending on the water content of the hair. We achieve a fast and stable simulation by implementing the method based on position based dynamics for both hair and liquid. As a result, we confirmed that our method can broaden the range of hair expressions and makes it possible to simulate the hair deformation due to the interaction with liquid.

ABSTRACT. Vapor condensation is a commonly observed phenomenon in everyday life. However, it has not reached a sufficient level because of its complicated transition of appearance. In this paper, we propose a technique for rendering vapor deposition using model-based techniques. Based on our observation that the vapor deposition's appearance can be explained mostly by heat propagation and water-repellency of the surface of an object, our method controls the appearance by the parameters obtained from them. In the experiment, the transition of vapor deposition on complicated shapes with different materials was synthesized, which confirms the effectiveness of the proposed method.

ABSTRACT. Thin-film interference is widely known for the reason of the iridescence in the bubble's surface and spilled oil. In the real world environment, oil films happen more naturally and are seen more often than soap bubbles. It also has more variations depending on what materials it is attached to. Calculating the interference of the light waves is not expensive, but in real-time rendering, the calculation is still considered impractical to use. Calculating interference in advance and use it as a texture could be a solution but inflexible and has low accuracy due to the limited texture size. With this in mind, we propose a lightweight method to render thin-film interference that won't hinder the real-time rendering process by using a 1-dimensional interference map with optical path difference as an index. With 1 dimensional map, the interference can be saved with a higher range and accuracy in small size.

ABSTRACT. We propose a system for rendering the night sky in real-time on a web browser. Our system builds on the WebGL technology, which has been widely used for displaying 3D contents on the web. However, we find that WebGL is too slow for computing the intensities of the sky based on a physically-based model. We also need to use a time-consuming tone mapping operator to create final images to be displayed. We thus combine the WebAssembly technology with WebGL to achieve higher performance. Since these are cross-platform technologies, the user does not need to prepare any special environments. We demonstrate that our system can render realistic images of the night sky in about 15 fps using a standard laptop computer.