In this paper, we propose a kansei retrieval method for living room images so that users can intuitively retrieve their desired properties on real estate portal sites. The proposed method employs a two-stages training model that combines Conv-DCCAE (Deep Canonically Correlated Auto-Encoders) and ranking model. For training model, a set of anchor images, positive examples of kansei words that represent impressions of the anchor images, and negative examples of kansei words that differ from the impression of the anchor images are input. First, the relationship between kansei words and images is trained using Conv-DCCAE, and projected onto the common space. Then, a ranking model is trained, and a kansei retrieval space is constructed by correcting the distance of positive examples to the anchors to be smaller and the distance of negative examples to the anchors to be larger in the common space. The proposed method also personalize the kansei retrieval space by retraining the ranking model using relevance feedback. The retrieval performance was evaluated by using nDCG (normalized Discounted Cumulative Gain), which is an index to evaluate the performance of the ranking. The proposed method achieves the most accurate ranking for nDCG@1, 5, 10 compared with the conventional methods. In addition, the proposed method was able to provide more appropriate retrieval results to users, whose initial retrieval results were inappropriate, by using relevance feedback.

Isaka et al. developed a naked-eye stereoscopic viewing system by inducing motion parallax through the orthogonal arrangement of two general-purpose display monitors and combining anamorphosis, which is known as a monocular illusion expression, with tracking of the user’s line of sight. However, the stereoscopic effect perceived with this system may possibly be degraded when using small display monitors due to the effect of binocular disparity. In this paper, we aim at pinpointing the effect of anamorphosis by identifying the difference in the contribution of the two eyes in recognizing the composition of the displayed space. We empirically proved that rather than tracking the dominant eye, it is possible to pinpoint the effect and to improve the sense of presence by placing “cyclopean eye” at a certain internal position between the two eyes to obtain proper anamorphosis. An example of porting the proposed method to a commodity laptop PC with a foldable display is also presented.

Brittle fracture of plane shape objects, such as glass and concrete, is often seen in the real world. Fracture animation of rigid bodies provides impressive effects by using physics-oriented simulation. However, simulation costs become too high when physics-oriented simulation approaches are chosen to generate realistic fracture animation. On the other hand, pre-fractured patterns with Voronoi-based segmentation applied when a collision occurs are usually used in real-time applications such as games. The geometry of such patterns is however monotonous and is hard to represent complicated fracture shapes realistically. There is thus a lack of trade-off methods that can realize realistic features as well as low computational costs. In this paper, we propose an alternative machine learning scheme based on conditional Generative Adversarial Network (cGAN) techniques to replace Voronoi-based segmentation for plane shape objects, which can be applied to rigid body engines such as Bullet Physics. Our learning datasets are generated by Boundary Element Method (BEM) based-fracture simulation. Compared with Voronoi-based segmentation, our method can generate much more complicated fracture shapes realistically at reasonable computational costs.

In this paper, we propose a method to reduce the computational complexity of the s Small Steps Method in Projective Dynamics, which is one of the elastic body simulation methods. The trade-off between the quality of the simulation and the amount of computation is always an issue in interactive computer graphics scenes. The Small Steps Method improves the simulation by reducing the numerical dumping by dividing the time evolution between drawing frames into small sub-steps instead of terminating the iterative computation for solving the problem in the middle. We note that the small-step method implicitly assumes the computation of sub-steps that are not drawn, and we omit the redundancy in the time direction. Specifically, we introduce DummyStep, a forward Eulerian method that uses acceleration to perform time evolution, and substitute DummyStep for half of the substeps in Projective Dynamics. As a result, Dummy Steps Method can replace the Small Steps Method with a very small error, and the computational complexity can be reduced by about half. However, since Dummy Steps Method has some stability problems, we propose two modification methods to further improve the stability of the method. Through example problems, we demonstrate that our method can stably replace the small step method while significantly reducing the computational complexity.

Digital holography is a technology that records three-dimensional (3D) information as a hologram using an image sensor and visualizes it on a computer. It is possible to obtain the amplitude and phase of the sample by calculation. However, the resolution of current image sensors is not sufficient, and in-line holography using plane waves is often used to obtain higherdefinition reproduced images. However, in-line holography has the disadvantage that unnecessary optical information overlaps with the desired 3D information. In addition, the optical system becomes large, such as the need for a collimator lens to create a plane wave. Therefore, we propose a system that applies optimization calculation to in-line digital holography using a reference light of spherical waves, which can simplify the optical system rather than plane waves. Furthermore, we found a new estimation method that does not require iterative calculation. In numerical experiments, the proposed method is 50 times faster than the commonly used Gerchberg-Saxton method for the same image quality.

In this study, we propose a method to transform projections into a sharp model based on parameters calculated from point cloud model. In the conventional editing process of point cloud models, the models are often converted to curved surface models or polygon models. However, if the point cloud data has noise or scan error, it is difficult to generate correct point connections and surfaces. In the proposed method, the point cloud data is directly processed. Specifically, parameters are calculated for each point using point cloud morphology operations, and the model is sharpened by moving the points using the parameters. The effectiveness of the proposed method is verified through experiments of parameter calculation when the size of the structural elements is changed and sharp transformation.

In the invisible digital watermarking method, image quality, resistance, and embedding information amount are in a trade-off relationship with each other, and it is usually difficult to satisfy all of them. The Green-Noise Diffused Watermarking Method has print/scan resistance and invisible by human eye. However, the amount of embedding information depends on the block size, and if the block size is reduced to increase the amount of embedded information, false judgement will increase. Therefore, we introduced machine learning to improve the identification accuracy. As a result, embedding 2048 bits in a 512 × 512 image with a 16 × 16 block size, we obtained correct answer rate of 99% for electronic data and 91% for printed/scanned images of 200ppi (94% for images of 175ppi or less), and achieved enhancing both robustness and embedding amount of information.

One of the simplest way to realize a stereoscopic style transfer is using monocular style transfer to stereo image as normal still image. However, deformation caused by style transfer deteriorates the disparity information of stereo image, so improvement of style transfer needs a evaluation method for transferred image. In this paper, evaluation image of disparity information for stereoscopic style transfer of stereo image and its numerical evaluation method are proposed. Proposed evaluation image determines three regions from transferred image, they are ”correct disparity region, loss of disparity region, incorrect disparity”. And numerical evaluation in each region employs their area whose value is normalized. In an evaluation experiment, parameter of evaluation image is considered, and the basic property of the them are showed. Finally, its effectiveness for stereoscopic style transfer is also showed by using pencil art generation method.

In this project, we are developing a system to remove objects, logos, annotations, and noises from videos with as little human intervention as possible by using SiamMask and a video completion method, both of which are existing methods. The user specifies a target object by drawing a bounding box around it in the first frame of the video; this bounding box is taken as input by SiamMask. SiamMask then tracks the target object and produces its mask in each frame. The resulting masks are then taken as input by a video completion method, which produces the final video completion result. The goal of this project is that, after drawing the bounding box, the user immediately obtains the video completion result. However, the mask produced by our current method is not always perfect. When imperfections arise, the user still has to manually modify the mask using an image editing software. As an application of this method, we also propose a method to generate a highly accurate mask of the target object based on the difference between the input video and the restoration result.

A framework for activity detection in surveillance videos generally involves activity proposal generation and activity classification. An activity proposal is a spatial and temporal candidate region for an arbitrary activity, and an activity classifier identifies the activity class for activity proposals. One of the difficulties in activity classification is the increase in the number of appearance patterns due to the diversity of the object moving directions and object-positional relationships. To solve this problem, we propose an activity alignment method for rotating activity proposals so that the object-movement direction and object-positional relationship are constant among all activity proposals before classification. The experimental results indicate that activity detection accuracy improves by adding our method to a baseline method using the ActEV/VIRAT dataset.