No description available.

The work described in this thesis contributes to the development of a system to evaluate sailing performance. This work was motivated by the lack of tools available to evaluate sailing performance. The goal of the work presented is to detect and classify the turns of a sailing yacht. Data was collected using a BlackBerry PlayBook affixed to a J/24 sailing yacht. This data was manually annotated with three types of turn: tack, gybe, and mark rounding. This manually annotated data was used to train classification methods. Classification methods tested were multi-layer perceptrons (MLPs) of two sizes in various committees and nearest- neighbour search. Pre-processing algorithms tested were Kalman filtering, categorization using quantiles, and residual normalization. The best solution was found to be an averaged answer committee of small MLPs, with Kalman filtering and residual normalization performed on the input as pre-processing.

In this thesis, the problem of online adaptive hashing for fast similarity search is studied. Similarity search is a central problem in many computer vision applications. The ever-growing size of available data collections and the increasing usage of high-dimensional representations in describing data have increased the computational cost of performing similarity search, requiring search strategies that can explore such collections in an efficient and effective manner. One promising family of approaches is based on hashing, in which the goal is to map the data into the Hamming space where fast search mechanisms exist, while preserving the original neighborhood structure of the data. We first present a novel online hashing algorithm in which the hash mapping is updated in an iterative manner with streaming data. Being online, our method is amenable to variations of the data. Moreover, our formulation is orders of magnitude faster to train than state-of-the-art hashing solutions. Secondly, we propose an online supervised hashing framework in which the goal is to map data associated with similar labels to nearby binary representations. For this purpose, we utilize Error Correcting Output Codes (ECOCs) and consider an online boosting formulation in learning the hash mapping. Our formulation does not require any prior assumptions on the label space and is well-suited for expanding datasets that have new label inclusions. We also introduce a flexible framework that allows us to reduce hash table entry updates. This is critical, especially when frequent updates may occur as the hash table grows larger and larger. Thirdly, we propose a novel mutual information measure to efficiently infer the quality of a hash mapping and retrieval performance. This measure has lower complexity than standard retrieval metrics. With this measure, we first address a key challenge in online hashing that has often been ignored: the binary representations of the data must be recomputed to keep pace with updates to the hash mapping. Based on our novel mutual information measure, we propose an efficient quality measure for hash functions, and use it to determine when to update the hash table. Next, we show that this mutual information criterion can be used as an objective in learning hash functions, using gradient-based optimization. Experiments on image retrieval benchmarks confirm the effectiveness of our formulation, both in reducing hash table recomputations and in learning high-quality hash functions.

Made available in DSpace on 2014-06-12T15:57:17Z (GMT). No. of bitstreams: 2 arquivo3157_1.pdf: 3737373 bytes, checksum: 7ca491f9a72f2e9cf51764a7acac3e3c (MD5) license.txt: 1748 bytes, checksum: 8a4605be74aa9ea9d79846c1fba20a33 (MD5) Previous issue date: 2010 / Conselho Nacional de Desenvolvimento Científico e Tecnológico / Esta dissertação introduz uma estrutura de dados baseada em gride implementada em GPU. Ela foi desenvolvida para pesquisa dos vizinhos mais próximos em nuvens de pontos dinâmicas, de uma forma massivamente paralela. A implementação possui desempenho em tempo real e é executada em GPU, ambas construção do gride e pesquisas dos vizinhos mais próximos (exatos e aproximados). Dessa forma, a transferência de memória entre sistema e dispositivo é minimizada, aumentando o desempenho de uma forma geral. O algoritmo proposto pode ser usado em diferentes aplicações com cenários estáticos ou dinâmicos. Além disso, a estrutura de dados suporta nuvens de pontos tridimensionais e dada sua natureza dinâmica, o usuário pode mudar seus parâmetros em tempo de execução. O mesmo se aplica ao número de vizinhos pesquisados. Uma referência em CPU foi implementada e comparações de desempenho justificam o uso de GPUs como processadores massivamente paralelos. Em adição, o desempenho da estrutura de dados proposta é comparada com implementações em CPU e GPU de trabalhos anteriores. Finalmente, uma aplicação de renderização baseada em pontos foi desenvolvida de forma a verificar o potencial da estrutura de dados

The work described in this thesis contributes to the development of a system to evaluate sailing performance. This work was motivated by the lack of tools available to evaluate sailing performance. The goal of the work presented is to detect and classify the turns of a sailing yacht. Data was collected using a BlackBerry PlayBook affixed to a J/24 sailing yacht. This data was manually annotated with three types of turn: tack, gybe, and mark rounding. This manually annotated data was used to train classification methods. Classification methods tested were multi-layer perceptrons (MLPs) of two sizes in various committees and nearest- neighbour search. Pre-processing algorithms tested were Kalman filtering, categorization using quantiles, and residual normalization. The best solution was found to be an averaged answer committee of small MLPs, with Kalman filtering and residual normalization performed on the input as pre-processing.

Word embeddings are useful in many tasks in Natural Language Processing and Information Retrieval, such as text mining and classification, sentiment analysis, sentence completion, or dictionary construction. Word2vec and its predecessor fastText, both well-known models to produce word embeddings, are powerful techniques to study the syntactic and semantic relations between words by representing them in a low-dimensional vector. By applying algebraic operations on these vectors semantic relationships such as word analogies, gender-inflections, or geographical relationships can be easily recovered. The aim of this work is to investigate how word embeddings could be utilized to augment and enrich queries in DBMSs, e.g. to compare text values according to their semantic relation or to group rows according to the similarity of their text values. For this purpose, we use pre-trained word embedding models of large text corpora such as Wikipedia. By exploiting this external knowledge during query processing we are able to apply inductive reasoning on text values. Thereby, we reduce the demand for explicit knowledge in database systems. In the context of the IMDB database schema, this allows for example to query movies that are semantically close to genres such as historical fiction or road movie without maintaining this information. Another example query is sketched in Listing 1, that returns the top-3 nearest neighbors (NN) of each movie in IMDB. Given the movie “Godfather” as input this results in “Scarface”, “Goodfellas” and “Untouchables”.

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第21390号 / 情博第676号 / 新制||情||117(附属図書館) / 京都大学大学院情報学研究科知能情報学専攻 / (主査)教授 鹿島 久嗣, 教授 山本 章博, 教授 下平 英寿 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

This large body of work is entirely centered around dual-tree algorithms, a class of algorithm based on spatial indexing structures that often provide large amounts of acceleration for various problems. This work focuses on understanding dual-tree algorithms using a new, tree-independent abstraction, and using this abstraction to develop new algorithms. Stated more clearly, the thesis of this entire work is that we may improve and expand the class of dual-tree algorithms by focusing on and providing improvements for each of the three independent components of a dual-tree algorithm: the type of space tree, the type of pruning dual-tree traversal, and the problem-specific BaseCase() and Score() functions. This is demonstrated by expressing many existing dual-tree algorithms in the tree-independent framework, and focusing on improving each of these three pieces. The result is a formidable set of generic components that can be used to assemble dual-tree algorithms, including faster traversals, improved tree theory, and new algorithms to solve the problems of max-kernel search and k-means clustering.

Local feature detection and its applications in different problems are quite popular in vision research. In order to analyze a scene, its invariant features, which are distinguishable in many views of this scene, are used in pose estimation, object detection and augmented reality. However, required performance metrics might change according to the application type / in general, the main metrics are accepted as accuracy and computational complexity. The contributions in this thesis provide improving these metrics and can be divided into three parts, as local feature detection, local feature description and description matching in different views of the same scene. In this thesis an efficient feature detection algorithm with sufficient repeatability performance is proposed. This detection method is convenient for real-time applications. For local description, a novel local binary pattern outperforming state-of-the-art binary pattern is proposed. As a final task, a fuzzy decision tree method is presented for approximate nearest neighbor search. In all parts of the system, computational efficiency is considered and the algorithms are designed according to limited processing time. Finally, an overall system capable of matching different views of the same scene has been proposed and executed in a mobile platform. The results are quite promising such that the presented system can be used in real-time applications, such as augmented reality, object retrieval, object tracking and pose estimation.

Nearest-neighbor search is a very natural and universal problem in computer science. Often times, the problem size necessitates approximation. In this thesis, I present new paradigms for nearest-neighbor search (along with new algorithms and theory in these paradigms) that make nearest-neighbor search more usable and accurate. First, I consider a new notion of search error, the rank error, for an approximate neighbor candidate. Rank error corresponds to the number of possible candidates which are better than the approximate neighbor candidate. I motivate this notion of error and present new efficient algorithms that return approximate neighbors with rank error no more than a user specified amount. Then I focus on approximate search in a scenario where the user does not specify the tolerable search error (error constraint); instead the user specifies the amount of time available for search (time constraint). After differentiating between these two scenarios, I present some simple algorithms for time constrained search with provable performance guarantees. I use this theory to motivate a new space-partitioning data structure, the max-margin tree, for improved search performance in the time constrained setting. Finally, I consider the scenario where we do not require our objects to have an explicit fixed-length representation (vector data). This allows us to search with a large class of objects which include images, documents, graphs, strings, time series and natural language. For nearest-neighbor search in this general setting, I present a provably fast novel exact search algorithm. I also discuss the empirical performance of all the presented algorithms on real data.