Columbia Object Image Library (COIL-20) is a database of gray-scale images of 20 objects. The objects were placed on a motorized turntable against a black background. The turntable was rotated through 360 degrees to vary object pose with respect to a fixed camera. Images of the objects were taken at pose intervals of 5 degrees. This corresponds to 72 images per object. The database has two sets of images. The first set contains 720 unprocessed images of 10 objects. The second contains 1,440 size normalized images of 20 objects. COIL-20 is available online via ftp. i 1 Introduction We have constructed a database of 1,440 grayscale images of 20 objects (72 images per object). The objects have a wide variety of complex geometric and reflectance characteristics (see figure 1(a)). The database, called Columbia Object Image Library (COIL-20), was used in a real-time 20 object recognition system [ Murase and Nayar-1995 ] . Figure 1(b) shows an object from the database being placed in front...

Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE. Contact: Manager,

A system to retrieve images using a syntactic description of appearance is presented. A multiscale invariant vector representation is obtained by first filtering images in the database with Gaussian derivative filters at several scales and then computing low order differential invariants. The multi-scale representation is indexed for rapid retrieval. Queries are designed by the users from an example image by selecting appropriate regions. The invariant vectors corresponding to these regions are matched with those in the database both in feature space as well as in coordinate space and a match score is obtained for each image. The results are then displayed to the user sorted by the match score. From experiments conducted with over 1500 images it is shown that images similar in appearance and whose viewpoint is within 25 degrees of the query image can be retrieved with an average precision of 57.4% 1. INTRODUCTION The goal of image retrieval systems is to operate on collections of imag...

Recently, we have proposed a new approach to estimation of the coefficients of eigenimages, which is robust against occlusion, varying background, and other types of non-Gaussian noise [4, 5]. In this paper we show that our method for estimating the coefficients can be applied to convolved and subsampled images yielding the same value of the coefficients. This enables an efficient multiresolution approach, where the values of the coefficients can directly be propagated through the scales. This property is used to extend our robust method to the problem of scaled images. We performed extensive experimental evaluations to confirm our theoretical results. 1 Introduction The appearance-based approaches to vision problems based on eigenspace analysis have shown the potential in many successful applications [10, 11, 15, 8, 14, 1]. The main advantage of these approaches is that the models encompass 2-D views which can easily be learned and enable to deal with combined effects of shape, refl...

In this paper, we examine the use of machine learning to improve the robustness of systems for image analysis on the task of roof detection. We review the problem of analyzing aerial photographs, and describe an existing vision system that attempts to automate the identification of buildings in aerial images. After this, we briefly review several well-known learning algorithms that represent a wide variety of inductive biases. We report three experiments designed to illuminate facets of applying machine learning methods to the image analysis task; one experiment focuses on within-image learning, another deals with the cost of different errors, and a third addresses between-image learning. Experimental results demonstrate that machinelearned classifiers meet or exceed the accuracy of handcrafted solutions and that useful generalization occurs when training and testing on data derived from different images. 1 Introduction The number of images available to image analysts is growing rapid...

The retrieval of images based on their visual similarity to an example image is an important and fascinating area of research. Here, a method to characterize visual appearance for determining global similarity in images is described. Images are filtered with Gaussian derivatives and geometric features are computed from the filtered images. The geometric features used here are curvature and phase. Two images may be said to be similar if they have similar distributions of such features. Global similarity may, therefore, be deduced by comparing histograms of these features. This allows for rapid retrieval and examples from collection of gray-level and trademark images are shown. 1 Introduction The advent of large multi-media collections and digital libraries has led to a need for good search tools to index and retrieve information from them. For text available in machine readable form (ASCII) a number of good search engines are available. However, there are as yet no good tools to retri...

The retrieval of images based on their visual similarity to an example image is an important and fascinating area of research. Here, we discuss various ways in which visual appearance may be characterized for determining both global and local similarity in images. One popular method involves the computation of global measures like moment invariants to characterize global similarity. Although this means that the image may be characterized using a few numbers, the performance is often poor. Techniques based on moment invariants often perform poorly. They require that the object be a binary shape without holes which is often not practical. In addition, moment invariants are sensitive to noise. Visual appearance is better represented using local features computed at multiple scales. Such local features may include the outputs of images filtered with Gaussian derivatives, differential invariants or geometric quantities like curvature and phase. Two images may be said to be similar if they have similar distributions of such features. Global similarity may, therefore, be deduced by comparing histograms of such features. This can be done rapidly. Histograms cannot be used to compute local similarity. Instead, the constraint that the spatial relationship between the features in the query be similar to the spatial relationship between the features of its matching counterparts in the database provides a means for computing local similarity. The methods presented here do not require prior segmentation of the database. In the case of local representation objects can be embedded in arbitrary backgrounds and both methods handle a range of size variations and viewpoint variations up to 20 or 25 degrees. Keywords: filter based representations, appearance based representations, scale ...

TV commercials recognition is a need for advertisers in order to check the fulfillment of their contracts with TV stations. In this paper we presentanapproachtothis problem based on compacting a representative frame of each shot by a PCA of its color histogram. Wealso present a new algorithm for scene break detection based on the analysis of local color variations in consecutive frames of some specific regions of the image. 1

For many computer vision applications labeled/segmented data is needed. Manually assigning labels or segmenting images is a time consuming and tedious task and becomes infeasible for a huge amount of data (e.g., when analyzing a video stream). Thus, this paper proposes a new approach to minimize the manual labeling/segmentation effort for learning an object detector by automatically extracting training data directly from a video sequence. Therefore, a robust background model, a tracker and an on-line learning method are combined. The main idea is to track an object through a video sequence and to directly use the obtained image patches, showing the object from different views, to incrementally update an existing model which in turn can be used for detection. As the tracker is initialized automatically by change detection, no user interaction is needed! Thus, an unknown object can be learned without having any prior information. To show the benefit of the proposed approach the framework is demonstrated on several typical objects that can be found on a desktop. 1

The goal of image retrieval is to retrieve images "similar" to a given query image by comparing the query and database using visual attributes like color, texture and appearance. In this paper, we discuss how to characterize appearance and use it for image retrieval. Visual appearance is represented by the outputs of a set of Gaussian derivative filters applied to an image. These outputs are computed off-line and stored in a database. A query is created by outlining portions of the query image deemed useful for retrieval by the user (this may be changed interactively depending on the results). The query is also filtered with Gaussian derivatives and these outputs are compared with those from the database. The images in the database are ranked on the basis of this comparison. The technique has been experimentally tested on a database of 1600 images which includes a variety of images. The system does not require prior segmentation of the database. Objects can be embedded in arbitrary backgrounds. The system handles a range of size variations and viewpoint variations up to 20 or 25 degrees. Keywords: filter based representations, appearance based representations, scale space matching, vector correlation, rmage retrieval, image indexing. 1.