| CHENG ZHUO | M, 1回目発表 | 生体医用画像 | 佐藤 嘉伸, | 向川 康博, | 大竹 義人, | Soufi Mazen, | 上村圭亮 |
|
Title: Fully Automatic Vertebra Segmentation for Spine and Pelvis Alignment Analysis in a Large-scale CT Database Abstract: Our project aims at the analysis of musculoskeletal structures for preoperative surgical planning and prediction of postoperative changes in orthopaedic surgery, specifically in the hip region. Analysis of the pelvis and spine alignment is important and has a potential for a variety of clinical applications, while the multi-class segmentation of each vertebrae poses a number of challenges including similarity between the neighboring vertebrae, large variation of the field of view (FOV) and the range of scans in the protocols commonly used in clinical practice. In this research, we aim to propose a deep learning based fully automatic and robust vertebra segmentation method for alignment analysis in a large-scale CT database. Language of the presentation: Japanese | |||||||
| HU XIAODAN | M, 1回目発表 | サイバネティクス・リアリティ工学 | 清川 清, | 向川 康博, | 酒田 信親, | 磯山 直也 | |
|
title: Soft-edge Occlusion Glasses for Autism Spectrum Disorder
abstract: It is common for people with Autism Spectrum Disorder (ASD) to present problems with atypical visual perception. Due to neuronal disabilities, specific visual stimulation like strong ambient light or fast movement objects will lead to abnormal visual symptoms. We propose a novel design for smart glasses with automatic occlusion to darken the ambient light that comes from bright regions which will reduce the visual stimulation brought to the user. A transparent LCD is used as the glasses lens, while a scene camera is responsible for stimulation detection. In this opportunity, I will be presenting the actual state of the project, the advantages, and deficiencies of the current prototype and the future plans. language of the presentation: English | |||||||
| CHANG XIN | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 向川 康博, | 諏訪 博彦, | 松田 裕貴 | |
|
title: Collecting Tourists' Activity from Public Cameras Using Open-World Person Re-identification
abstract: Mainstream travel recommender systems perform trip recommendation by solving a tourist trip design problem (TTDP), which greatly relies on the quantity and quality of statistical data (visiting time, travelling cost, tourist satisfaction, etc.). However, collecting these data requires active user participation that is both high-cost and time-consuming. In this research, we leverage person re-identification (re-ID) to construct a vision-based system for automatic acquisition of tourists' visiting activity at points of interest (PoIs). Application of re-ID methods into an open-world setting faces two main challenges: 1) difficulty in maintaining a dynamic ID gallery and 2) heavy computation burden of global queries. We propose to introduce hierarchical clustering from offline multi-object tracking algorithms to efficiently reduce the redundancy in re-ID gallery and limit the computation time. Meanwhile, the imbalance of intra-ID feature distribution is also relieved. A narrower gallery scale and querying scope enabled our data-collection system to operate in real-time in a near-online manner. Our system will be evaluated in terms of re-ID average precision, average coverage of re-identified time and visiting time estimation error. language of the presentation: English | |||||||
| DANG CHENYU | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一, | 清川 清, | 諏訪 博彦, | 松田 裕貴 | |
|
title: Real-Time Taxi Demand Hotspots Detection Using Computer Vision Approach
abstract: Unbalanced taxi demand and supply always cause not only passengers to wait for a long time, but also the taxi company to rise their operation cost. To address this, many researchers are focusing on predicting taxi demand hotspots as precise as possible with various types of data(e.g., weather, events, etc.). Most of them can only get the predicted result in a rough area. On the other hand, predicted results not always precise due to the changeful situation. We proposed this system to improve this problem. In general, Our system operates in a two-step paradigm: firstly, we process the streaming video data from taxi's drive recorder to analyze the status(e.g., waving hand.) of pedestrians through computer vision technologies (Object Detection, Multi-Object Tracking and Action Recognition). Subsequently, we correct the analized result with the location data by implementing a Random Forest regressor. Because the visual information can most intuitively reflect the situation of streets, our approach provides a more direct solution to find taxi demand hotspots in the real-time. language of the presentation: English | |||||||
| FU XINGJIAN | M, 2回目発表 | 光メディアインタフェース | 向川 康博, | 清川 清, | 舩冨 卓哉, | 田中 賢一郎, | 久保 尋之 |
|
title: A flexible method for single shot camera distortion rectification
abstract: Single shot camera distortion has long been a problem of optimization. Recently, deep learning based distortion rectification have shown great potential as an alternative to traditional approaches. For simplicity, many of the existing deep learning frameworks assume one camera model of fixed focal length, and the nature of distortions are usually expressed as a model specific set of parameters. This poses great disadvantages when dealing with real world optical systems, in which distortions are usually more complicated than anticipated. To extend deep learning’s capability in predicting distortions, we take several distortion models into consideration and propose a versatile neural network. The models of choice are integrated into one by employing more intuitive distortion maps. The results indicate that our model is capable of predicting the underlying distortion maps in both synthetic and real world images. language of the presentation: English | |||||||
| 山野 仁詩 | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, | 宮尾 知幸, | MD.Altaf-Ul-Amin, | 小野 直亮, | 黄 銘 |
|
title: Clustering various properties of polymers and estimating them based on monomer unit structure information
abstract: In polymer material development, multiple physical and chemical properties need to be simultaneously optimized. However, polymer properties are usually more difficult to predict than those of small molecules due to them forming superstructures. In this work, we aimed at finding a versatile approach to predict multiple polymer properties using imperfect data with missing values. We prepared a data set consisting of 50 polymers with 45 properties along with their monomer unit structures. The data set was hierarchically clustered on the basis of two independent factors: polymer properties and polymer structures. In polymer property-based clustering, visualizing relations of polymers was found to be an effective way of estimating the difficulty of polymer property prediction. In polymer structure-based clustering, each cluster could be formed based on the structural features. Thus, the clustering contributed to understanding structural characteristics of monomer unit structures. In addition to analyzing the data set in an unsupervised manner, we constructed polymer properties prediction models based solely on the information of monomer unit structures. Using Dragon7, which is commercial descriptor generation software, constitutional and topological descriptors were obtained for the monomer unit structures. Partial least squared regression (PLSR) models could predict density, glass transition temperature, and dissolution parameter with high accuracy. Values of R2 for test data sets were between 0.88 and 0.97. language of the presentation: Japanese | |||||||