This paper proposes a stress-based elastic body fracture simulation method suitable for parallel computing using GPUs. By integrating three techniques: Position Based Dynamics for simulating elastic bodies, a meshless stress calculation method, and a crack propagation and fracture surface generation method using surfels and crack tip models, dynamic and flexible fracture surface generation can be efficiently processed. This resolves the trade-off relationship between computational cost and crack shape representation ability that conventional fracture simulations have. In addition, the proposed method is implemented using Compute Shader, and experiments using GPUs confirm that it is possible to achieve both detailed fracture surface representation and real-time processing.

Thin-film interference, as observed in soap bubbles, is an optical phenomenon caused by the nature of light waves. Spectral rendering and spectral reflectance models are required to accurately render the phenomenon in computer graphics (CG). In particular, it is important to be able to handle various materials in rendering, and several spectral reflectance models that take into account surface roughness have already been proposed. However, existing models do not consider multiple scattering between the micro-surfaces, resulting in energy loss on the rough surfaces. In this paper, we propose a spectral reflectance model of thin-film interference that considers multiple scattering. The proposed method assumes that the surface consists of v-groove micro-surfaces, and both the thin-film interference and the multiple scattering are taken into account by integrating the micro-facet reflection model that considers multiple scattering and the multilayer model that considers light interference and absorption. We demonstrate that the proposed method can accurately render the interference effects on rough surfaces, especially when the surface roughness increases.

Three-dimensional (3D) point set, or point cloud, is a 3D shape representation that can be captured by a 3D sensor or derived from a CAD model. Accurate analysis of 3D point sets is required for effective reuse of 3D models, or navigation and control of autonomous vehicles and robots. These 3D point sets are in general not aligned rotationally. But many applications of these 3D point sets require robustness against rotation of 3D shapes. Recent 3D point set analysis relies on Deep Neural Networks (DNN), yet most of these DNN are not robust against rotation. In this paper, we propose and evaluate a rotation invariant 3D shape analysis DNN. The DNN combines rotation normalization of local geometry using local reference frame with content adaptive feature extraction via self-attention mechanism. We evaluate the DNN on both shape classification and segmentation of 3D point sets. The proposed method is invariant to rotation about 3 axes while outperforming existing methods in accuracy.

The functionalities of smartphone applications or IoT technologies have been widely enhanced, and deep learning models have taken a role in many recent products. Also, the need to adopt deep learning models in edge devices is increasing. However, it is challenging to load models with enormous parameters on edge devices because of memory or power shortage. Model reduction by parameter approximation enables large models on small devices and enlarges application areas. This paper proposes leveraging model reduction using matrix factorization with well-known parameter reduction methods, pruning, and quantization. Since the target of matrix factorization is the weight matrix of linear calculation, we apply our proposed method to one of the recent popular models in the image recognition area, a Transformer model, most parameters of which belongs to its fully connected layers. Experimental results show that our method can reduce the size of the trained CIFAR10 model up to 60% of that of the original model. We can use the ratio of the sign of output vector elements in each layer as the index of layer selection as the objective of pruning.

In this paper, we propose a method where we represent the 3D shape of mackerels and horse mackerels flowing on a conveyor as point clouds and recognize fish species using PointNet, a point cloud data classification model. Specifically, when fish flowing on a conveyor are photographed from above(Z-axis), the rotation of the fish is assumed to be only around the Z-axis, and we propose a PointNet that is limited to rotation normalization around the Z-axis only. We conducted experiments using point cloud data which are photographed at a fish market, and the following were revealed. (1) We investigated the relations between the number of points used for recognition, accuracy rate and the recognition speed, and our findings showed that using 128 points resulted in the best recognition performance. (2) Comparing the accuracy rate of the proposed method with that of the original PointNet, the proposed method gave a higher accuracy rate for both 100 stationary fish and 308 fish moved by a conveyor. (3)Comparison of the proposed method with the method that recognizes fish species from color and depth images acquired simultaneously with the point cloud using the ResNet-50 image classification model showed that the accuracy rate of the proposed method using the point cloud is the same or higher than that of the method using color and depth images, respectively.

In writer identification with application in such as handwriting analysis, grasp of intra/inter-variation (variability and individuality) is the most important problem. This paper proposes a simple measurement method of intra/inter-writer variation on each part in handwriting, and clearly presents intra/inter-writer variation along strokes in kanji 6 characters, hiragana 5 characters, katakana 5 characters, and numeric 10 characters. Furthermore, this paper proposes extraction methods of significant parts in handwriting for writer identification, and demonstrate experimental results about those 26 characters.