RESEARCH OUTPUTS

  • JOURNAL ARTICLE
    A Survey of 3D Reconstruction with Event Cameras
    1 Oct 2026Computational Visual Media12(5):1263-1299Tsinghua University Press
    Event cameras are rapidly emerging as powerful vision sensors for 3D reconstruction, uniquely capable of asynchronously capturing per-pixel brightness changes. Compared to traditional frame-based cameras, event cameras produce sparse yet temporally dense data streams, enabling robust and accurate 3D reconstruction even under challenging conditions such as high-speed motion, low illumination, and extreme dynamic range. These capabilities offer substantial promise for transformative applications in various fields, including autonomous driving, robotics, aerial navigation, and immersive virtual reality. This survey presents the first comprehensive review exclusively dedicated to event-based 3D reconstruction. Existing approaches are systematically categorized based on input modality as stereo, monocular, or multimodal systems, and further classified according to reconstruction methodologies, including geometry-based techniques, deep learning approaches, and neural rendering techniques such as neural radiance fields (NeRFs) and 3D Gaussian splatting (3DGS). Within each category, methods are chronologically organized to highlight the evolution of key concepts and advances. We also provide a detailed summary of publicly available datasets specifically suited to event-based reconstruction tasks. Finally, we discuss significant open challenges in dataset availability, standardized evaluation, effective representation, and dynamic scene reconstruction, suggesting directions for future research. This survey aims to serve as a reference and to provide a clear and motivating roadmap for advancing the state-of-the-art in event-driven 3D reconstruction.
  • JOURNAL ARTICLE
    Automatic Construction of Fine-Grained Paraphrase Corpora System Using Language Inference Model
    2022Applied Sciences12(1):499-1MDPIAG
    Co-authors:
    • Zhou Y,
    • Hu X,
    • Chung VY
    DOIDOI: 10.3390/app12010499 (opens in a new tab)⁠
    Paraphrase detection and generation are important natural language processing (NLP) tasks. Yet the term paraphrase is broad enough to include many fine-grained relations. This leads to different tolerance levels of semantic divergence in the positive paraphrase class among publicly available paraphrase datasets. Such variation can affect the generalisability of paraphrase classification models. It may also impact the predictability of paraphrase generation models. This paper presents a new model which can use few corpora of fine-grained paraphrase relations to construct automatically using language inference models. The fine-grained sentence level paraphrase relations are defined based on word and phrase level counterparts. We demonstrate that the fine-grained labels from our proposed system can make it possible to generate paraphrases at desirable semantic level. The new labels could also contribute to general sentence embedding techniques.
  • CONFERENCE
    On the Usability of Hadoop MapReduce, Apache Spark & Apache Flink for Data Science
    1 Dec 20172017 IEEE International Conference on Big Data (Big Data)303-310Institute of Electrical and Electronics Engineers (IEEE)
    Distributed data processing platforms for cloud computing are important tools for large-scale data analytics. Apache Hadoop MapReduce has become the de facto standard in this space, though its programming interface is relatively low-level, requiring many implementation steps even for simple analysis tasks. This has led to the development of more advanced dataflow oriented platforms, most prominently Apache Spark and Apache Flink. Those platforms not only aim to improve performance through improved in-memory processing, but in particular provide built-in high-level data processing functionality, such as filtering and join operators, which should make data analysis tasks easier to develop than with plain Hadoop MapReduce. But is this indeed the case? This paper compares three prominent distributed data processing platforms: Apache Hadoop MapReduce; Apache Spark; and Apache Flink, from a usability perspective. We report on the design, execution and results of a usability study with a cohort of master students, who were learning and working with all three platforms in order to solve different use cases set in a data science context. Our findings show that Spark and Flink are preferred platforms over MapReduce. Among participants, there was no significant difference in perceived preference or development time between both Spark and Flink as platforms for batch-oriented big data analysis. This study starts an exploration of the factors that make Big Data platforms more-or less-effective for users in data science.
  • CONFERENCE
    Effectiveness of Adversarial Attacks on Class-Imbalanced Convolutional Neural Networks
    2017Neural Information ProcessingLecture Notes in Computer Science10635:333-342Springer Nature
    Convolutional neural networks (CNNs) performance has increased considerably in the last couple of years. However, as with most machine learning methods, these networks suffer from the data imbalance problem - when the underlying training dataset is comprised of an unequal number of samples for each label/class. Such imbalance enforces a phenomena known as domain shift that causes the model to have poor generalisation when presented with previously unseen data. Recent research has focused on a technique called gradient sign that intensifies domain shift in CNNs by modifying inputs to deliberately yield erroneous model outputs, while appearing unmodified to human observers. Several commercial systems rely on image recognition techniques to perform well. Therefore, adversarial attacks poses serious threats to their integrity. In this work we present an experimental study that sheds light on the link between adversarial attacks, imbalanced learning and transfer learning. Through a series of experiments we evaluate the fast gradient sign method on class imbalanced CNNs, linking model vulnerabilities to the characteristics of its underlying training set and internal model knowledge.
  • CONFERENCE
    Multi-Swarms Dynamic Convergence Optimization for Object Tracking
    1 Jul 20162016 International Joint Conference on Neural Networks (IJCNN)155-162Institute of Electrical and Electronics Engineers (IEEE)
    Co-authors:
    • Sha F,
    • Huang W,
    • Chung YY
    DOIDOI: 10.1109/ijcnn.2016.7727193 (opens in a new tab)⁠
    Swarm intelligence has been applied to many research projects in recent years, many scientists are working on developing the full potential of a self-organized and decentralized system to help solving complex problems. In image processing, it also demonstrates fast and accurate in searching solutions for trajectory clustering and precise object tracking. This paper is aim to introduce a novel multiple particle swarms with dynamic convergence approach for object tracking in complicated environment. Our new approach absorbs the advantages of other multi swarm algorithms to optimize the resources and process iteration. So it can provide more accurate and faster tracking result for both linear and non-linear movement pattern when compared to basic PSO and other PSO based algorithms such as inertia weight PSO and constriction factor PSO. In addition, multiple independent populations will not only inherit each of their own attribute's weights through dynamic range convergence, but also influence by each other's solution effects. The experiments have been conducted with different types of testing videos in real environment. The results examined with different types of moving pattern have demonstrated that the new method required less resources and iteration process and could have better tracking performance and scarcely lost target with diverse interferences.
  • CONFERENCE
    Profiling Apache HIVE Query from Run Time Logs
    1 Jan 20162016 International Conference on Big Data and Smart Computing (BigComp)61-68Institute of Electrical and Electronics Engineers (IEEE)
    Apache Hive is a widely used data warehousing and analysis tool. Developers write SQL like HIVE queries, which are converted into MapReduce programs to runs on a cluster. Despite its popularity, there is little research on performance comparison and diagnose. Part of the reason is that instrumentation techniques used to monitor execution can not be applied to intermediate MapReduce code generated from Hive query. Because the generated MapReduce code is hidden from developers, run time logs are the only places a developer can get a glimpse of the actual execution. Having an automatic tool to extract information and to generate report from logs is essential to understand the query execution behavior. We designed a tool to build the execution profile of individual Hive queries by extracting information from HIVE and Hadoop logs. The profile consists of detailed information about MapRe-jobs, tasks and attempts belonging to a query. It is stored as a JSON document in MongoDB and can be retrieved to generate reports in charts or tables. We have run several experiments on AWS with TPC-H data sets and queries to demonstrate that our profiling tool is able to assist developers in comparing HIVE queries written in different formats, running on different data sets and configured with different parameters. It is also able to compare tasks/attempts within the same job to diagnose performance issues.
  • CONFERENCE
    Systematic Analysis of Public Cloud Service Level Agreements and Related Business Values
    28 Oct 20132013 IEEE International Conference on Services Computing729-736Institute of Electrical and Electronics Engineers (IEEE)
    Co-authors:
    • Qiu MM,
    • Zhou Y,
    • Wang C
    DOIDOI: 10.1109/scc.2013.24 (opens in a new tab)⁠
    Cloud computing offers lots of attractive benefits to individuals and organizations. However, the adoption rate is a far cry from what it really deserves. Lack of well defined Service Level Agreement (SLA) is one of the key factors. This research aims to provide better understanding of the status quo of the current public cloud service level agreements and to recommend best practices. We conducted a content analysis on around thirty SLAs to identify commonly stated attributes and commonly missing attributes. A follow up case study interviewed a selection of cloud customers further investigates the importance of SLA attributes and its relationship with potential business. The result reveals that some of the commonly missed attributes such as customer support, data protection plan are considered as very important attributes and most SLA attributes are directly related with business values.
  • CONFERENCE
    An Online Query Authentication System for Outsourced Databases
    1 Jul 20132013 12th IEEE International Conference on Trust, Security and Privacy in Computing and Communications570-577Institute of Electrical and Electronics Engineers (IEEE)
    The cloud computing paradigm moves increasing amount of data across administrative domains to be managed by various outsourced database (ODB) service providers. There needs a means for a data owner to verify the correctness and completeness of query results returned from these ODB services. The process is called query authentication. This paper proposes to use independent observers for query authentication. This approach enables the data owner to maintain the data state using these observers and the query results can be verified by these observers without any additional information from the ODB. To preserve privacy to a certain degree, the data state is partitioned vertically among observers, therefore no observer has full knowledge about the data. We implement the method on top of a NoSQL database. Our evaluation results in Amazon EC2 show the effectiveness of the method in such kind of databases.
  • CONFERENCE
    A Query Verification Method for Making Outsourced Databases Trustworthy
    1 Jun 20122012 IEEE Ninth International Conference on Services Computing298-305Institute of Electrical and Electronics Engineers (IEEE)
    The cloud computing paradigm moves data across administrative domains to be managed by various outsourced database (ODB) service providers. There needs a means for a data owner to verify the correctness and completeness of query results returned from these ODB services. The process is called query authentication. Existing ODB services do not provide such assurance to customers. This is mainly due to that most techniques in literature are expensive and limited in functionality. This paper proposes to use independent external observers for query result verification. The proposed method requires the ODB service to return a verification object $({\cal VO})$ for each query result set, as most existing methods do. Differing to existing methods, we propose a complement set based ${\cal VO}$ construction method. The cost of computing complement set based ${\cal VO}$ is low compared to hash tree based ones commonly seen in existing methods. The proposed method does not require the data owner or the ODB service provider to maintain complex authentication data structures, instead, the workload of data state tracking and ${\cal VO}$ validation can be distributed among a set of observers. Furthermore, the proposed ${\cal VO}$ construction method is capable of supporting multi-attribute queries. The method is implemented as a thin wrapper of an ODB service with the support of additional observer services. We evaluate the performance of the system in Amazon EC2 and examine its cost of achieving trustworthiness.
  • CONFERENCE
    Combining content and quality indicators in ranking ambiguous query results on flickr
    1 Jan 2012Conferences in Research and Practice in Information Technology Series124:109-116
    Co-authors:
    • Zhong H,
    • Zhou Y
  • JOURNAL ARTICLE
    A decentralized search engine for dynamic Web communities
    Jan 2011Knowledge and Information Systems26(1):105-125Springer Nature
    Currently, most Web search engines perform search on corpus comprising nearly entire content of the Web. The same centralized search service can be performed on a single site as well. Nonetheless, there is little research on community-wide search. This paper presents a peer-to-peer search engine ComSearch. ComSearch is designed to provide small- and middle-scale online communities—the ability to perform text search within the community. Communities are formed in a self-organizing style. P2P IR system may suffer unnecessary internal traffic in answering a multi-term query. In this paper, we propose several techniques to optimize the multi-term query process. The simulation results show that our proposed algorithms have good scalability. Compared with baseline approach, our improved algorithm can reduce the communication cost by about two orders of magnitude in the best case. We also deploy the system in a small-scale network and conduct a series of experiments to estimate the actual query response time as well as to investigate the data movement effect caused by node joining. Experimental results show that multiple data movements are quite common during network expansion. However, the percentage of multiple data movements decreases when a network is getting stable after the initial frequent joining activities. This provides possibilities for improvement on P2P data movement management.
  • CONFERENCE
    Adding context to social tagging systems
    1 Dec 2010Acis 2010 Proceedings 21st Australasian Conference on Information Systems
    Co-authors:
    • Song JH,
    • Zhou Y,
    • Jung H
  • CONFERENCE
    Ontological services using crowdsourcing
    1 Dec 2010Acis 2010 Proceedings 21st Australasian Conference on Information Systems
    Co-authors:
  • CONFERENCE
    An Asymmetric Similarity Measure for Tag Clustering on Flickr
    1 Apr 20102010 12th International Asia-Pacific Web Conference171-177Institute of Electrical and Electronics Engineers (IEEE)
    Web 2.0 tools and environments have made tagging, the act of assigning keywords to on-line objects, a popular way to annotate shared resources. The success of now-prominent tagging systems makes tagging “the natural way for people to classify objects as well as an attractive way to discover new material”. One of the most challenging problems is to harvest the semantics from these systems, which can support a number of applications, including tag clustering and tag recommendation. We conduct detailed studies on different types of tag relations and tag similarity measures, and propose a scalable measure that we name Reliability Factor Similarity Measure (RFSM). We compare it with two other measures having similar scalability by integrating them into hierarchical clustering methods and performing tag clustering on a subset of Flickr data. The results suggest that RFSM outperforms those two measures when it is applies for tag clustering purpose. We also present an alternative way of utilizing discovered tag relations to set up tag refining rules in order to deal with some noise in the initial tag sets, which can in turn improve the precision of tag relations.
  • CONFERENCE
    A collaborative monitoring mechanism for making a multitenant platform accountable
    1 Jan 20102nd Usenix Workshop on Hot Topics in Cloud Computing Hotcloud 2010
    Co-authors:
    • Wang C,
    • Zhou Y
  • CONFERENCE
    Category Recommendation in User Specified Structure
    2010E-Commerce and Web TechnologiesLecture Notes in Business Information Processing61:24-35Springer Nature
    Tagging has become a main tool for Internet users to describe and advertise various web resources. The relatively flat structure of the tag space poses lots of challenges in tag based query engines. Many data-centric algorithms have been proposed to discover structures from the flat tag space and to improve query results. At the same time, lots of social networking sites start to provide mechanisms allowing users to specify simple hierarchical structures. The group concept in Flickr is a good example of such user specified structure. Users can create a group with predefined themes. Other users can add their resources to several related groups voluntarily or by invitation. These groups are analogue to categories in a cataloguing system. This user specified structure would ideally improve the precision of tag based query. However, categories can be created by any user and public categories like groups are open for users to add resources in. More often than not, a simple category title does not give enough information on its content. In this paper we propose two algorithms, traditional IR cosine similarity approach and frequent pattern matching approach, to recommend categories to a given resource. We evaluate our algorithms using groups and photos from Flickr. Both algorithms achieve promising results in terms of precision in general. We also analyse strength and weakness of the two algorithms with respect to features of test data. We believe such recommendation mechanism is an important complement to any user specified hierarchical structure.
  • CONFERENCE
    An Integrated Approach to Extracting Ontological Structures from Folksonomies
    2009The Semantic Web: Research and ApplicationsLecture Notes in Computer Science5554:654-668Springer Nature
    Collaborative tagging systems have recently emerged as one of the rapidly growing web 2.0 applications. The informal social classification structure in these systems, also known as folksonomy, provides a convenient way to annotate resources by allowing users to use any keyword or tag that they find relevant. In turn, the flat and non-hierarchical structure with unsupervised vocabularies leads to low search precision and poor resource navigation and retrieval. This drawback has created the need for ontological structures which provide shared vocabularies and semantic relations for translating and integrating the different sources. In this paper, we propose an integrated approach for extracting ontological structure from folksonomies that exploits the power of low support association rule mining supplemented by an upper ontology such as WordNet.
  • CONFERENCE
    Integration of Computational and Crowd-Sourcing Methods for Ontology Extraction
    1 Jan 20092009 Fifth International Conference on Semantics, Knowledge and Grid306-309Institute of Electrical and Electronics Engineers (IEEE)
    Ontological approach is one of those promising solutions to improve search precision, resource navigation and retrieval by providing shared vocabularies and semantic relations for translating and integrating the different sources. But the progress has slowed down due to the knowledge acquisition bottleneck. In this paper, we present an integrated conceptual framework for extracting ontological structures from folksonomies that exploits the combined power of traditional computation and pooled, human knowledge through crowd-sourcing.
  • CONFERENCE
    Searching and Clustering on Social Tagging Sites
    1 Jan 20092009 Fifth International Conference on Semantics, Knowledge and Grid99-105Institute of Electrical and Electronics Engineers (IEEE)
    Social Tagging Site has increasingly become a main avenue for people to share resources online. Users can use simple tools to publish everything from bookmarks to video clips on those sites. However, effective querying of resources in such sites is still a challenging question in industry and academic research. This paper reports a novel algorithm of presenting the web resources query result on the social tagging sites. It adopts a two step clustering approach to organize and rank resources based on their relative similarities with each other. Initial term similarities are computed using user and tag information of the resource. The query results are then organized as a group of concepts represented by a few semantically related terms. The resources that are related with each concept are rank with respect to the concept. In addition, concepts are also ranked by representing terms and the number of resources associated.
  • CONFERENCE
    Query optimizing on a decentralized web search engine
    11 Mar 2007Proceedings of the 2007 ACM symposium on Applied computing878-879Association for Computing Machinery (ACM)
    Currently, most web search engines perform search on nearly a whole copy of the web corpus. There are also tools to add search functionality to a single site. Nonetheless, there is little research into search related to online communities: set of related websites or weblogs. To fill this gap, a peer-to-peer search engine is presented in this paper. This p2p search engine is designed to provide small- and middle-scale online communities such as blog sites and news sites the ability to perform text search within the community. It organizes nodes on a Distributed Hash Table (DHT) based peer-to-peer network. Communities are formed in a self-organizing style. P2P IR systems may cause increased internal traffic among nodes in answering a multi-term query. In this paper, we focused on this issue and proposed several techniques to optimize the multi-term query process in a P2P framework. Our proposed algorithms are evaluated by simulation. The simulation results show that our proposed algorithms have good scalability and can improve performance of the system by about two orders of magnitude in the best case.
  • CONFERENCE
    Analysis of weblog link structure - A community perspective
    1 Dec 2006Webist 2006 2nd International Conference on Web Information Systems and Technologies ProceedingsSEBEG(EL/-):13-20
    Co-authors:
  • CONFERENCE
    Automatic extraction of complex web data
    1 Dec 2006Pacis 2006 10th Pacific Asia Conference on Information Systems ICT and Innovation Economy1436-1449
    Co-authors:
    • Zhang M,
    • Zhou Y,
    • Patrick J
  • CONFERENCE
    Community discovery and analysis in blogspace
    23 May 2006Proceedings of the 15th international conference on World Wide Web1017-1018Association for Computing Machinery (ACM)
    Weblog has quickly evolved into a new information and knowledge dissemination channel. Yet it is not easy to discover weblog communities through keyword search. The main contribution of this paper is the study of weblog communities from the perspective of social network analysis. We proposed a new way of collecting and preparing data for weblog community discovery. The data collection stage focuses on gaining knowledge of the strength of social ties between weblogs. The strength of social ties and the clustering feature of social network guided the discovery of weblog communities.
  • CONFERENCE
    A Self-Organizing Search Engine for RSS Syndicated Web Contents
    1 Jan 200622nd International Conference on Data Engineering Workshops (ICDEW'06)1-9Institute of Electrical and Electronics Engineers (IEEE)
    The exponentially growing information published on the Web relies largely on a few major search engines like Google to be brought to the public nowadays. This raises issues such as: 1. how many percents of coverage do these search engines provide for the whole shared contents over the Internet? 2. how easy is it to find less popular contents from the Web through the page ranking system of these search engines? In fact, the increasing dynamics of the information distributed on the Internet challenge the flexibility of these centralized search engines. With the amount of structured and semi-structured data increase on the Internet, self-organizing search engines that are capable of providing sufficient coverage for data that follow certain structures get more and more attractive. In this paper, we propose a self-organizing search engine soSpace for RSS syndicated web data. soSpace is built on structured peer-to-peer technology. It enables indexing and searching of frequently updated web information described by RSS feed. Our experiment results show that it has good scalability as the contents increase. The recall and precision rate of the result are satisfactory as well.
  • CONFERENCE
    Open source software reliability model
    2005Proceedings of the fifth workshop on Open source software engineering - 5-WOSSE1-6Association for Computing Machinery (ACM)
    We collected bug tracking data from a few popular open source projects and investigated the time related bug reporting patterns from them. The results indicate that along its development cycle, open source projects exhibit similar reliability growth pattern with that of closed source project. Bug arrivals of most open source project will stabilize at a very low level, even though in comparison, no formal testing activities are involved. This stabilizing point would be viewed as the mature point for adoption consideration. The results also show that general Weibull distribution offers possible way to establish the reliability model; Also, popular measures such as page views and download are not highly correlated with the bug arrival rate and may not be suitable measures for a project's quality.