nearest-neighbors search, randomized kd-trees, hierarchical k-means tree, clustering. For many computer vision problems, the most time consuming component consists of nearest neighbor matching in high-dimensional spaces. There are no known exact algorithms for solving these high-dimensional problems that are faster than linear search. Approximate algorithms are known to provide large speedups with only minor loss in accuracy, but many such algorithms have been published with only minimal guidance on selecting an algorithm and its parameters for any given problem. In this paper, we describe a system that answers the question, “What is the fastest approximate nearest-neighbor algorithm for my data? ” Our system will take any given dataset and desired degree of precision and use these to automatically determine the best algorithm and parameter values. We also describe a new algorithm that applies priority search on hierarchical k-means trees, which we have found to provide the best known performance on many datasets. After testing a range of alternatives, we have found that multiple randomized k-d trees provide the best performance for other datasets. We are releasing public domain code that implements these approaches. This library provides about one order of magnitude improvement in query time over the best previously available software and provides fully automated parameter selection. 1

Abstract not found

The Internet contains billions of images, freely available online. Methods for efficiently searching this incredibly rich resource are vital for a large number of applications. These include object recognition [2], computer graphics [11, 27], personal photo collections, online image search tools. In this paper, our goal is to develop efficient image search and scene matching techniques that are not only fast, but also require very little memory, enabling their use on standard hardware or even on handheld devices. Our approach uses recently developed machine learning techniques to convert the Gist descriptor (a real valued vector that describes orientation energies at different scales and orientations within an image) to a compact binary code, with a few hundred bits per image. Using our scheme, it is possible to perform real-time searches with millions from the Internet using a single large PC and obtain recognition results comparable to the full descriptor. Using our codes on high quality labeled images from the LabelMe database gives surprisingly powerful recognition results using simple nearest neighbor techniques. Recent interest in object recognition has yielded a wide range of approaches to describing the contents of an image. One important application for this technology is the visual search of large collections of images, such as those on the Internet or on people’s home computers. Accordingly, a number of recognition papers have explored this area. Nister and Stewenius demonstrate the real-time specific object recognition using a database of 40,000 images [19]; Obdrzalek and Matas show sub-linear indexing time on the COIL dataset [20]. A common theme is the representation of the image as a collection of feature vectors and the use of efficient data structures to handle the large num-

This paper presents a MapReduce algorithm for computing pairwise document similarity in large document collections. MapReduce is an attractive framework because it allows us to decompose the inner products involved in computing document similarity into separate multiplication and summation stages in a way that is well matched to efficient disk access patterns across several machines. On a collection consisting of approximately 900,000 newswire articles, our algorithm exhibits linear growth in running time and space in terms of the number of documents. 1

Abstract. This paper introduces an approach for dense 3D reconstruction from unregistered Internet-scale photo collections with about 3 million images within the span of a day on a single PC (“cloudless”). Our method advances image clustering, stereo, stereo fusion and structure from motion to achieve high computational performance. We leverage geometric and appearance constraints to obtain a highly parallel implementation on modern graphics processors and multi-core architectures. This leads to two orders of magnitude higher performance on an order of magnitude larger dataset than competing state-of-the-art approaches. 1

Language Processing. Do not distribute! We propose an automatic method for measuring music similarity using audio features so we can enhance the current generation of taxonomy-based music search engines and recommender systems. Efficiency is important in an Internet-connected world, where users have access to millions of tracks. Brute-force algorithms for searching through this content are not practical. Many previous approaches to track similarity require pair-wise processing between all audio features in a database and therefore are generally not practical for large collections. Our features are time-ordered overlapping fixed-length subsequences of equal-temperament pitch-class profiles and log-frequency cepstral coefficients; the technique is analogous to the technique of shingling used for text retrieval. We use locality sensitive hashing to implement approximate matching for our high-dimensional audio shingles. This approach retrieves near neighbors within a specified distance of the query rather than retrieving only the nearest neighbors; the degree of approximation, ɛ, is a parameter. LSH achieves sub linear query time performance with respect to the number of tracks in a collection but requires an accurate threshold on retrieval distance for efficient performance. In this paper, we present a new method for estimating the optimal search radius for LSH retrieval tasks by modeling the between-shingle distance distributions for non-similar audio shingles. We derive an estimator for a minimum distance for two shingles to be considered drawn from different tracks. therefore, are considered to be drawn from similar tracks. We evaluate our proposed methods on three contrasting music similarity tasks: retrieval of mis-attributed recordings (Apocrypha), retrieval of the same work by performed by different artists (Opus) and retrieval of edited and sampled versions of a query track by remix artists (Remixes). Our results achieve near-perfect performance in the first two tasks and 80 % precision at 70 % recall in the third task.

This paper addresses the problem of designing binary codes for high-dimensional data such that vectors that are similar in the original space map to similar binary strings. We introduce a simple distribution-free encoding scheme based on random projections, such that the expected Hamming distance between the binary codes of two vectors is related to the value of a shift-invariant kernel (e.g., a Gaussian kernel) between the vectors. We present a full theoretical analysis of the convergence properties of the proposed scheme, and report favorable experimental performance as compared to a recent state-of-the-art method, spectral hashing. 1

Image classification is a critical task for both humans and computers. One of the challenges lies in the large scale of the semantic space. In particular, humans can recognize tens of thousands of object classes and scenes. No computer vision algorithm today has been tested at this scale. This paper presents a study of large scale categorization including a series of challenging experiments on classification with more than 10, 000 image classes. We find that a) computational issues become crucial in algorithm design; b) conventional wisdom from a couple of hundred image categories on relative performance of different classifiers does not necessarily hold when the number of categories increases; c) there is a surprisingly strong relationship between the structure of WordNet (developed for studying language) and the difficulty of visual categorization; d) classification can be improved by exploiting the semantic hierarchy. Toward the future goal of developing automatic vision algorithms to recognize tens of thousands or even millions of image categories, we make a series of observations and arguments about dataset scale, category density, and image hierarchy.

Motivated by our work with political scientists who need to manually analyze large Web archives of news sites, we present SpotSigs, a new algorithm for extracting and matching signatures for near duplicate detection in large Web crawls. Our spot signatures are designed to favor naturallanguage portions of Web pages over advertisements and navigational bars. The contributions of SpotSigs are twofold: 1) by combining stopword antecedents with short chains of adjacent content terms, we create robust document signatures with a natural ability to filter out noisy components of Web pages that would otherwise distract pure n-gram-based approaches such as Shingling; 2) we provide an exact and efficient, selftuning matching algorithm that exploits a novel combination of collection partitioning and inverted index pruning for high-dimensional similarity search. Experiments confirm an increase in combined precision and recall of more than 24 percent over state-of-the-art approaches such as Shingling or I-Match and up to a factor of 3 faster execution times than Locality Sensitive Hashing (LSH), over a demonstrative “Gold Set ” of manually assessed near-duplicate news articles as well as the TREC WT10g Web collection.

We say that an algorithm for nearest neighbor search is combinatorial if only direct comparisons between two pairwise similarity values are allowed. Combinatorial algorithms for nearest neighbor search have two important advantages: (1) they do not map similarity values to artificial distance values and do not use the triangle inequality for the latter, and (2) they work for arbitrarily complicated data representations and similarity functions. In this paper we introduce a special property of the similarity function on a set S that leads to efficient combinatorial algorithms for S. The disorder constant D(S) of a set S is defined to ensure the following inequality: if x is the a’th most similar object to z and y is the b’th most similar object to z, then x is among the D(S) · (a + b) most similar objects to y. Assuming that disorder is small we present the first two known combinatorial algorithms for nearest neighbors whose query time has logarithmic dependence on the size of S. The first one, called Ranwalk, is a randomized zero-error algorithm that always returns the exact nearest neighbor. It uses space quadratic in the input size in preprocessing, but is very efficient in query processing. The second algorithm, called Arwalk, uses near-linear space. It uses random choices in preprocessing, but the query processing is essentially deterministic. For an arbitrary query q, there is only a small probability that the chosen data structure does not support q. Finally, we show that for the Reuters corpus average disorder is indeed quite small and that Ranwalk efficiently computes the nearest neighbor in most cases.