DOCSLIB.ORG

Enhancing Multi-Lingual Information Extraction Via Cross-Media Inference and Fusion

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Enhancing Multi-Lingual Information Extraction Via Cross-Media Inference and Fusion Enhancing Multi-lingual Information Extraction via Cross-Media Inference and Fusion Adam Lee, Marissa Passantino, Heng Ji Guojun Qi, Thomas Huang Computer Science Department Department of Electrical and Computer Queens College and Graduate Center Engineering & Beckman Institute City University of New York University of Illinois at Urbana-Champaign [email protected] [email protected] Abstract The processing methods of texts and images/videos are typically organized into two We describe a new information fusion separate pipelines. Each pipeline has been approach to integrate facts extracted studied separately and quite intensively over the from cross-media objects (videos and past decade. It is critical to move away from texts) into a coherent common represen- single media processing, and instead toward tation including multi-level knowledge methods that make multiple decisions jointly (concepts, relations and events). Beyond using cross-media inference. For example, video standard information fusion, we ex- analysis allows us to find both entities and ploited video extraction results and sig- events in videos, but it’s very challenging to nificantly improved text Information Ex- specify some fine-grained semantic types such traction. We further extended our meth- as proper names (e.g. “Obama Barack”) and ods to multi-lingual environment (Eng- relations among concepts; while the speech lish, Arabic and Chinese) by presenting embedded and the texts surrounding these a case study on cross-lingual comparable videos can significantly enrich such analysis. On corpora acquisition based on video com- the other hand, image/video features can parison. enhance text extraction. For example, entity gender detection from speech recognition output 1 Introduction is challenging because of entity mention recognition errors. However, gender detection An enormous amount of information is widely from corresponding images and videos can available in various data modalities (e.g. speech, achieve above 90% accuracy (Baluja and Rowley, text, image and video). For example, a Web 2006). In this paper, we present a case study on news page about “Health Care Reform in gender detection to demonstrate how text and America” is composed with texts describing video extractions can boost each other. some events (e.g., Final Senate vote for the We can further extend the benefit of cross- reform plans, Obama signs the reform media inference to cross-lingual information agreement), images (e.g., images about various extraction (CLIE). Hakkani-Tur et al. (2007) government involvements over decades) and found that CLIE performed notably worse than videos/speech (e.g. Obama’s speech video about monolingual IE, and indicated that a major the decisions) containing additional information cause was the low quality of machine translation regarding the real extent of the events or (MT). Current statistical MT methods require providing evidence corroborating the text part. large and manually aligned parallel corpora as These cross-media objects exist in redundant input for each language pair of interest. Some and complementary structures, and therefore it recent work (e.g. Munteanu and Marcu, 2005; Ji, is beneficial to fuse information from various 2009) found that MT can benefit from multi- data modalities. The goal of our paper is to lingual comparable corpora (Cheung and Fung, investigate this task from both mono-lingual and 2004), but it is time-consuming to identify pairs cross-lingual perspectives. of comparable texts; especially when there is lack of parallel information such as news release include coreferred persons, geo-political entities dates and topics. However, the images/videos (GPE), locations, organizations, facilities, vehi- embedded in the same documents can provide cles and weapons; relations include 18 types additional clues for similarity computation (e.g. “a town some 50 miles south of Salzburg” because they are ‘language-independent’. We indicates a located relation.); events include the will show how a video-based comparison 33 distinct event types defined in ACE 2005 approach can reliably build large comparable (e.g. “Barry Diller on Wednesday quit as chief text corpora for three languages: English, of Vivendi Universal Entertainment.” indicates a Chinese and Arabic. “personnel-start” event). Names are identified and classified using an HMM-based name tag- 2 Baseline Systems ger. Nominals are identified using a maximum entropy-based chunker and then semantically We apply the following state-of-the-art text and classified using statistics from ACE training video information extraction systems as our corpora. Relation extraction and event extraction baselines. Each system can produce reliable are also based on maximum entropy models, confidence values based on statistical models. incorporating diverse lexical, syntactic, semantic 2.1 Video Concept Extraction and ontological knowledge. The video concept extraction system was 3 Mono-lingual Information Fusion developed by IBM for the TREC Video and Inference Retrieval Evaluation (TRECVID-2005) (Naphade et al., 2005). This system can extract 3.1 Mono-lingual System Overview 2617 concepts defined by TRECVID, such as "Hospital", "Airplane" and "Female-Person". It Multi-media uses support vector machines to learn the Document mapping between low level features extracted from visual modality as well as from transcripts and production related meta-features. It also ASR Speech Videos/Images exploits a Correlative Multi-label Learner (Qi et al., 2007), a Multi-Layer Multi-Instance Kernel Texts (Gu et al., 2007) and Label Propagation through Partitioning Linear Neighborhoods (Wang et al., 2006) to extract all other high-level features. For each Text Information Video Concept classifier, different models are trained on a set Extraction Extraction of different modalities (e.g., the color moments, wavelet textures, and edge histograms), and the Entities/ predictions made by these classifiers are Concepts Relations/Events combined together with a hierarchical linearly- weighted fusion strategy across different modalities and classifiers. Multi-level Concept Fusion 2.2 Text Information Extraction Global Inference We use a state-of-the-art IE system (Ji and Grishman, 2008) developed for the Automatic Enhanced Content Extraction (ACE) program1 to process Concepts/Entities texts and automatic speech recognition output. Relations/Events The pipeline includes name tagging, nominal mention tagging, coreference resolution, time Figure 1. Mono-lingual Cross-Media expression extraction and normalization, rela- Information Fusion and Inference Pipeline tion extraction and event extraction. Entities Figure 1 depicts the general procedure of our mono-lingual information fusion and inference 1 http://www.nist.gov/speech/tests/ace/ approach. After we apply two baseline systems (sequential keyframes) and background speech to the multi-media documents, we use a novel to align concepts. During this fusion process, we multi-level concept fusion approach to extract a compare the normalized confidence values pro- common knowledge representation across texts duced from two pipelines to resolve the follow- and videos (section 3.2), and then apply a global ing three types of cases: inference approach to enhance fusion results • Contradiction – A video fact contradicts a (section 3.3). text fact; we only keep the fact with higher confidence. 3.2 Cross-media Information Fusion • Redundancy – A video fact conveys the • Concept Mapping same content as (or entails, or is entailed by) a text fact; we only keep the unique parts of For each input video, we apply automatic speech the facts. recognition to obtain background texts. Then we • Complementary – A video fact and a text use the baseline IE systems described in section fact are complementary; we merge these 2 to extract concepts from texts and videos. We two to form more complete fact sets. construct mappings on the overlapped facts across TRECVID and ACE. For example, • A Common Representation “LOC.Water-Body” in ACE is mapped to In order to effectively extract compact informa- “Beach, Lakes, Oceans, River, River_Bank” in tion from large amounts of heterogeneous data, TRECVID. we design an integrated XML format to repre- Due to different characteristics of video clips sent the facts extracted from the above multi- and texts, these two tasks have quite different level fusion. We can view this representation as granularities and focus. For example, a set of directed “information graphs” G={Gi “PER.Individual” in ACE is an open set includ- (Vi, Ei)}, where Vi is the collection of concepts ing arbitrary names, while TRECVID only cov- from both texts and videos, and Ei is the collec- ers some famous proper names such as tion of edges linking one concept to the other, “Hu_Jintao” and “John_Edwards”. Geopolitical labeled by relation or event attributes. An exam- entities appear very rarely in TRECVID because ple is presented in Figure 2. This common rep- they are more explicitly presented in back- resentation is applied in both mono-lingual and ground texts. On the other hand, TRECVID de- multi-lingual information fusion tasks described fined much more fine-grained nominals than in next sections. ACE, for example, “FAC.Building-Grounds” in ACE can be divided into 52 possible concept Amina Abbas types such as “Conference_Buildings” and Spouse “Golf_Course” because they can be more easily Leader Child Yasser PLO Mahmoud detected based on video features. We also notice Abbas Abbas that
Load more

Recommended publications
  • IAMG Newsletter IAMG Newsletter
    No. 101 December 2020 IAMGIAMG NewsletterNewsletter Official Newsletter of the International Association for Mathematical Geosciences Contents ow, what a strange year! WYet with international Submit nominationS for 2021 GeorGeS matheron Lectur- travel stopped, conferences er and 2022 iamG diStinGuiShed Lecturer ................. 1 postponed, and campuses diStinGuiShed Lecturer updateS ....................................... 1 closed we’ve adapted. As can preSident’S forum .............................................................. 3 be expected, we’re a little light on news. Hopefully 2021 is a much better year worldwide! In member newS ...................................................................... 4 the meantime, hopefully you enjoy the cartoons below. paSt preSident prof Jenny mcKinLey eLection to the iuGS I’d like to welcome Peter Dowd and our new IAMG council and executive counciL ......................................................... 4 thank Jenny McKinley and a our outgoing council for all their SprinGer encycLopedia of mathematicaL GeoScienceS .. 4 hard work over the last four years. prince SuLtan bin abduLaziz internationaL Katie Silversides water prize .................................................................... 4 <> ieee GeoScience and remote SenSinG Society (GrSS) diS- tinGuiShed Lecturer (dL) ............................................. 4 diverSity and incLuSion in GeoScience ............................. 4 Student newS ...................................................................... 5 nancy Student
    [Show full text]
  • 25Th International Joint Conference on Artif Icial Intelligence New York
    25th International Joint Conference on Artificial Intelligence New York City, July 9–15, 2016 www.ijcai-16.org Organizing Institutions IJCAI AAAI The International Joint Conferences on Artificial Intelligence The Association for the Advancement of Artificial Intelligence IJCAI-16 Welcome to IJCAI 2016 Welcome to the Twenty-fifth International The conference includes workshops, tuto- Joint Conference on Artificial Intelligence! rials, exhibitions, demonstrations, invited talks, and paper/poster presentations. On We are delighted to welcome you to New Friday there will be an Industry Day, with York, one of the most exciting cities of presentations from the top AI companies; the world, to take part in IJCAI, the leading and there will be an AI Festival, open to conference on the thrilling field of Artificial the public, consisting of the IJCAI award Intelligence. AI today has a tremendous winner’s talks. You will not want to miss out impact. It is in all the media and makes a on this highlight, so plan to stay at IJCAI-16 real difference. At IJCAI-16, you will have the until the very end. opportunity to meet some of the world’s leading AI researchers, to learn first-hand The conference (including workshops about their newest research results and and tutorials) takes place at the Hilton in developments, and to catch up with current midtown Manhattan. The Hilton is just a AI trends in science and industry. And, of ten-minute walk from Central Park, where course, IJCAI-16 will be the perfect forum the conference banquet will be held, for presenting your own achievements, and half a block from the world-famous both to specialists in your field, and to the Museum of Modern Art, where we will have AI world in general.
    [Show full text]
  • University of California San Diego
    UNIVERSITY OF CALIFORNIA SAN DIEGO Sparse Recovery and Representation Learning A dissertation submitted in partial satisfaction of the requirements for the degree Doctor of Philosophy in Mathematics by Jingwen Liang Committee in charge: Professor Rayan Saab, Chair Professor Jelena Bradic Professor Massimo Franceschetti Professor Philip E. Gill Professor Tara Javidi 2020 Copyright Jingwen Liang, 2020 All rights reserved. The dissertation of Jingwen Liang is approved, and it is ac- ceptable in quality and form for publication on microfilm and electronically: Chair University of California San Diego 2020 iii DEDICATION To my loving family, friends and my advisor. iv TABLE OF CONTENTS Signature Page . iii Dedication . iv Table of Contents . .v List of Figures . vii Acknowledgements . viii Vita .............................................x Abstract of the Dissertation . xi Chapter 1 Introduction and Background . .1 1.1 Compressed Sensing and low-rank matrix recovery with prior infor- mations . .1 1.2 Learning Dictionary with Fast Transforms . .3 1.3 Deep Generative Model Using Representation Learning Techniques5 1.4 Contributions . .6 Chapter 2 Signal Recovery with Prior Information . .8 2.1 Introduction . .8 2.1.1 Compressive Sensing . .9 2.1.2 Low-rank matrix recovery . 10 2.1.3 Prior Information for Compressive Sensing and Low-rank matrix recovery . 12 2.1.4 Related Work . 13 2.1.5 Contributions . 21 2.1.6 Overview . 22 2.2 Low-rank Matrices Recovery . 22 2.2.1 Problem Setting and Notation . 22 2.2.2 Null Space Property of Low-rank Matrix Recovery . 23 2.3 Low-rank Matrix Recovery with Prior Information . 31 2.3.1 Support of low rank matrices .
    [Show full text]
  • Vanderbiltuniversitymedicalcenter
    VanderbiltUniversityMedicalCenter Medical Center Medical Center School of Medicine Hospital and Clinic Vanderbilt University 2008/2009 Containing general information and courses of study for the 2008/2009 session corrected to 30 June 2008 Nashville The university reserves the right, through its established procedures, to modify the require- ments for admission and graduation and to change other rules, regulations, and provisions, including those stated in this bulletin and other publications, and to refuse admission to any student, or to require the withdrawal of a student if it is determined to be in the interest of the student or the university. All students, full- or part-time, who are enrolled in Vanderbilt courses are subject to the same policies. Policies concerning noncurricular matters and concerning withdrawal for medical or emo- tional reasons can be found in the Student Handbook, which is on the Vanderbilt Web site at www.vanderbilt.edu/student_handbook. NONDISCRIMINATION STATEMENT In compliance with federal law, including the provisions of Title IX of the Education Amend- ments of 1972, Title VI of the Civil Rights Act of 1964, Sections 503 and 504 of the Reha- bilitation Act of 1973, and the Americans with Disabilities Act of 1990, Vanderbilt University does not discriminate on the basis of race, sex, religion, color, national or ethnic origin, age, disability, or military service in its administration of educational policies, programs, or activ- ities; its admissions policies; scholarship and loan programs; athletic or other university- administered programs; or employment. In addition, the university does not discriminate on the basis of sexual orientation consistent with university non-discrimination policy.
    [Show full text]
  • 2017 Medford/Somerville Massachusetts
    161ST Commencement Tufts University Sunday, May 21, 2017 Medford/Somerville Massachusetts Commencement 2017 Commencement 2017 School of Arts and Sciences School of Engineering School of Medicine and Sackler School of Graduate Biomedical Sciences School of Dental Medicine The Fletcher School of Law and Diplomacy Cummings School of Veterinary Medicine The Gerald J. and Dorothy R. Friedman School of Nutrition Science and Policy Jonathan M. Tisch College of Civic Life #Tufts2017 commencement.tufts.edu Produced by Tufts Communications and Marketing 17-653. Printed on recycled paper. Table of Contents Welcome from the President 5 Overview of the Day 7 Graduation Ceremony Times and Locations 8 University Commencement 11 Dear Alma Mater 14 Tuftonia’s Day Academic Mace Academic Regalia Recipients of Honorary Degrees 15 School of Arts and Sciences 21 Graduate School of Arts and Sciences School of Engineering School of Medicine and Sackler School 65 of Graduate Biomedical Sciences Public Health and Professional 78 Degree Programs School of Dental Medicine 89 The Fletcher School of Law 101 and Diplomacy Cummings School of Veterinary Medicine 115 The Gerald J. and Dorothy R. Friedman 123 School of Nutrition Science and Policy COMMENCEMENT 2017 3 Welcome from the President This year marks the 161st Commencement exercises held at Tufts University. This is always the high point of the academic year, and we welcome all of you from around the world to campus for this joyous occasion—the culmination of our students’ intellectual and personal journeys. Today’s more than 2,500 graduates arrived at Tufts with diverse backgrounds and perspectives. They have followed rigorous courses of study on our four Massachusetts campuses while enriching the life of our academic community.
    [Show full text]
  • 2019 International Joint Conference on Neural Networks (IJCNN 2019)
    2019 International Joint Conference on Neural Networks (IJCNN 2019) Budapest, Hungary 14-19 July 2019 Pages 1-774 IEEE Catalog Number: CFP19IJS-POD ISBN: 978-1-7281-1986-1 1/8 Copyright © 2019 by the Institute of Electrical and Electronics Engineers, Inc. All Rights Reserved Copyright and Reprint Permissions: Abstracting is permitted with credit to the source. Libraries are permitted to photocopy beyond the limit of U.S. copyright law for private use of patrons those articles in this volume that carry a code at the bottom of the first page, provided the per-copy fee indicated in the code is paid through Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. For other copying, reprint or republication permission, write to IEEE Copyrights Manager, IEEE Service Center, 445 Hoes Lane, Piscataway, NJ 08854. All rights reserved. *** This is a print representation of what appears in the IEEE Digital Library. Some format issues inherent in the e-media version may also appear in this print version. IEEE Catalog Number: CFP19IJS-POD ISBN (Print-On-Demand): 978-1-7281-1986-1 ISBN (Online): 978-1-7281-1985-4 ISSN: 2161-4393 Additional Copies of This Publication Are Available From: Curran Associates, Inc 57 Morehouse Lane Red Hook, NY 12571 USA Phone: (845) 758-0400 Fax: (845) 758-2633 E-mail: [email protected] Web: www.proceedings.com TABLE OF CONTENTS COMPARISON OF PROBABILISTIC MODELS AND NEURAL NETWORKS ON PREDICTION OF HOME SENSOR EVENTS ............................................................................................................................................1 Flávia Dias Casagrande ; Jim Tørresen ; Evi Zouganeli CYSTOID FLUID COLOR MAP GENERATION IN OPTICAL COHERENCE TOMOGRAPHY IMAGES USING A DENSELY CONNECTED CONVOLUTIONAL NEURAL NETWORK .....................................9 Plácido L.
    [Show full text]
  • Multi-Instance Pose Networks: Rethinking Top-Down Pose Estimation
    Multi-Instance Pose Networks: Rethinking Top-Down Pose Estimation Rawal Khirodkar1∗ Visesh Chari2 Amit Agrawal2 Ambrish Tyagi2 1Carnegie Mellon University 2Amazon Lab 126 [email protected] fviseshc, aaagrawa, [email protected] Abstract A key assumption of top-down human pose estimation approaches is their expectation of having a single per- son/instance present in the input bounding box. This often leads to failures in crowded scenes with occlusions. We propose a novel solution to overcome the limitations of this fundamental assumption. Our Multi-Instance Pose Network (MIPNet) allows for predicting multiple 2D pose instances within a given bounding box. We introduce a Multi-Instance Modulation Block (MIMB) that can adaptively modulate channel-wise feature responses for each instance and is parameter efficient. We demonstrate the efficacy of our ap- proach by evaluating on COCO, CrowdPose, and OCHuman datasets. Specifically, we achieve 70:0 AP on CrowdPose and 42:5 AP on OCHuman test sets, a significant improve- ment of 2:4 AP and 6:5 AP over the prior art, respectively. When using ground truth bounding boxes for inference, MIP- Net achieves an improvement of 0:7 AP on COCO, 0:9 AP on CrowdPose, and 9:1 AP on OCHuman validation sets com- Figure 1: 2D pose estimation networks often fail in presence of pared to HRNet. Interestingly, when fewer, high confidence heavy occlusion. (Left) Bounding boxes corresponding to two per- bounding boxes are used, HRNet’s performance degrades sons. (Middle) For both bounding boxes, HRNet predicts the pose (by 5 AP) on OCHuman, whereas MIPNet maintains a rela- for the front person and misses the occluded person.
    [Show full text]
  • Embodied Question Answering
    2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2018) Salt Lake City, Utah, USA 18-22 June 2018 Pages 1-731 IEEE Catalog Number: CFP18003-POD ISBN: 978-1-5386-6421-6 1/13 Copyright © 2018 by the Institute of Electrical and Electronics Engineers, Inc. All Rights Reserved Copyright and Reprint Permissions: Abstracting is permitted with credit to the source. Libraries are permitted to photocopy beyond the limit of U.S. copyright law for private use of patrons those articles in this volume that carry a code at the bottom of the first page, provided the per-copy fee indicated in the code is paid through Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. For other copying, reprint or republication permission, write to IEEE Copyrights Manager, IEEE Service Center, 445 Hoes Lane, Piscataway, NJ 08854. All rights reserved. *** This is a print representation of what appears in the IEEE Digital Library. Some format issues inherent in the e-media version may also appear in this print version. IEEE Catalog Number: CFP18003-POD ISBN (Print-On-Demand): 978-1-5386-6421-6 ISBN (Online): 978-1-5386-6420-9 ISSN: 1063-6919 Additional Copies of This Publication Are Available From: Curran Associates, Inc 57 Morehouse Lane Red Hook, NY 12571 USA Phone: (845) 758-0400 Fax: (845) 758-2633 E-mail: [email protected] Web: www.proceedings.com 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition CVPR 2018 Table of Contents Message from the General and Program Chairs cxviii Organizing Committee and Area
    [Show full text]
  • Lecture Notes in Computer Science 9218
    Lecture Notes in Computer Science 9218 Commenced Publication in 1973 Founding and Former Series Editors: Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen Editorial Board David Hutchison Lancaster University, Lancaster, UK Takeo Kanade Carnegie Mellon University, Pittsburgh, PA, USA Josef Kittler University of Surrey, Guildford, UK Jon M. Kleinberg Cornell University, Ithaca, NY, USA Friedemann Mattern ETH Zurich, Zürich, Switzerland John C. Mitchell Stanford University, Stanford, CA, USA Moni Naor Weizmann Institute of Science, Rehovot, Israel C. Pandu Rangan Indian Institute of Technology, Madras, India Bernhard Steffen TU Dortmund University, Dortmund, Germany Demetri Terzopoulos University of California, Los Angeles, CA, USA Doug Tygar University of California, Berkeley, CA, USA Gerhard Weikum Max Planck Institute for Informatics, Saarbrücken, Germany More information about this series at http://www.springer.com/series/7412 Yu-Jin Zhang (Ed.) Image and Graphics 8th International Conference, ICIG 2015 Tianjin, China, August 13–16, 2015 Proceedings, Part II 123 Editor Yu-Jin Zhang Department of Electronic Engineering Tsinghua University Beijing China ISSN 0302-9743 ISSN 1611-3349 (electronic) Lecture Notes in Computer Science ISBN 978-3-319-21962-2 ISBN 978-3-319-21963-9 (eBook) DOI 10.1007/978-3-319-21963-9 Library of Congress Control Number: 2015944504 LNCS Sublibrary: SL6 – Image Processing, Computer Vision, Pattern Recognition, and Graphics Springer Cham Heidelberg New York Dordrecht London © Springer International Publishing Switzerland 2015 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed.
    [Show full text]
  • Multi-Hypothesis Pose Networks: Rethinking Top-Down Pose
    Multi-Instance Pose Networks: Rethinking Top-Down Pose Estimation Rawal Khirodkar1* Visesh Chari2 Amit Agrawal2 Ambrish Tyagi2 1Carnegie Mellon University 2Amazon Lab 126 [email protected] fviseshc, aaagrawa, [email protected] Abstract A key assumption of top-down human pose estimation approaches is their expectation of having a single per- son/instance present in the input bounding box. This often leads to failures in crowded scenes with occlusions. We propose a novel solution to overcome the limitations of this fundamental assumption. Our Multi-Instance Pose Network (MIPNet) allows for predicting multiple 2D pose instances within a given bounding box. We introduce a Multi-Instance Modulation Block (MIMB) that can adaptively modulate channel-wise feature responses for each instance and is parameter efficient. We demonstrate the efficacy of our ap- proach by evaluating on COCO, CrowdPose, and OCHuman datasets. Specifically, we achieve 70:0 AP on CrowdPose and 42:5 AP on OCHuman test sets, a significant improve- ment of 2:4 AP and 6:5 AP over the prior art, respectively. When using ground truth bounding boxes for inference, MIP- Net achieves an improvement of 0:7 AP on COCO, 0:9 AP on Figure 1: 2D pose estimation networks often fail in presence of CrowdPose, and 9:1 AP on OCHuman validation sets com- heavy occlusion. (Left) Bounding boxes corresponding to two per- pared to HRNet. Interestingly, when fewer, high confidence sons. (Middle) For both bounding boxes, HRNet predicts the pose bounding boxes are used, HRNet’s performance degrades for the front person and misses the occluded person.
    [Show full text]
  • Meshfreeflownet: a Physics-Constrained Deep Continuous Space-Time Super-Resolution Framework
    MeshfreeFlowNet: A Physics-Constrained Deep Continuous Space-Time Super-Resolution Framework Chiyu \Max" Jiang∗1, Soheil Esmaeilzadeh∗2, Kamyar Azizzadenesheli3, Karthik Kashinath4, Mustafa Mustafa4, Hamdi A. Tchelepi2, Philip Marcus1, Prabhat4, and Anima Anandkumar3,5 1 University of California, Berkeley, CA 94720, USA 2Stanford University, Stanford, CA 94305, USA 3California Institute of Technology, Pasadena, CA, 91125, USA 4Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA 5NVIDIA, Santa Clara, CA 95051, USA Abstract From a numerical perspective, resolving the wide range of spatio-temporal scales within such physical systems is We propose MeshfreeFlowNet, a novel deep learning- challenging since extremely small spatial and temporal nu- based super-resolution framework to generate continuous merical stencils would be required. In order to alleviate the (grid-free) spatio-temporal solutions from the low-resolution computational burden of fully resolving such a wide range inputs. While being computationally efficient, Mesh- of spatial and temporal scales, multiscale computational ap- freeFlowNet accurately recovers the fine-scale quantities proaches have been developed. For instance, in the subsur- of interest. MeshfreeFlowNet allows for: (i) the output face flow problem, the main idea of the multiscale approach to be sampled at all spatio-temporal resolutions, (ii) a set is to build a set of operators that map between the unknowns of Partial Differential Equation (PDE) constraints to be im- associated with the computational cells in a fine-grid and the posed, and (iii) training on fixed-size inputs on arbitrarily unknowns on a coarser grid. The operators are computed sized spatio-temporal domains owing to its fully convolu- numerically by solving localized flow problems.
    [Show full text]
  • Minimally Interactive Segmentation with Application to Human Placenta in Fetal MR Images
    Minimally Interactive Segmentation with Application to Human Placenta in Fetal MR Images Guotai Wang A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy of University College London. Centre for Medical Image Computing Department of Medical Physics and Biomedical Engineering University College London June 8, 2018 2 3 I, Guotai Wang, confirm that the work presented in this thesis is my own. Where information has been derived from other sources, I confirm that this has been indicated in the work. Abstract Placenta segmentation from fetal Magnetic Resonance (MR) images is important for fetal surgical planning. However, accurate segmentation results are difficult to achieve for automatic methods, due to sparse acquisition, inter-slice motion, and the widely varying position and shape of the placenta among pregnant women. Interactive meth- ods have been widely used to get more accurate and robust results. A good interactive segmentation method should achieve high accuracy, minimize user interactions with low variability among users, and be computationally fast. Exploiting recent advances in machine learning, I explore a family of new interactive methods for placenta seg- mentation from fetal MR images. I investigate the combination of user interactions with learning from a single image or a large set of images. For learning from a single image, I propose novel Online Random Forests to efficiently leverage user interactions for the segmentation of 2D and 3D fetal MR images. I also investigate co-segmentation of multiple volumes of the same patient with 4D Graph Cuts. For learning from a large set of images, I first propose a deep learning-based framework that combines user interactions with Convolutional Neural Networks (CNN) based on geodesic distance transforms to achieve accurate segmentation and good interactivity.
    [Show full text]