The problem of large-scale image search has been traditionally addressed with the bag-of-visual-words (BOV). In this article, we propose to use as an alternative the Fisher kernel framework. We first show why the Fisher representation is well-suited to the retrieval problem: it describes an image by what makes it different from other images. One drawback of the Fisher vector is that it is high-dimensional and, as opposed to the BOV, it is dense. The resulting memory and computational costs do not make Fisher vectors directly amenable to large-scale retrieval. Therefore, we compress Fisher vectors to reduce their memory footprint and speed-up the retrieval. We compare three binarization approaches: a simple approach devised for this representation and two standard compression techniques. We show on two publicly available datasets that compressed Fisher vectors perform very well using as little as a few hundreds of bits per image, and significantly better than a very recent compressed BOV approach.

Kernel machines rely on an implicit mapping of the data such that non-linear classification in the original space corresponds to linear classification in the new space. As kernel machines are difficult to scale to large training sets, it has been proposed to perform an explicit mapping of the data and to learn directly linear classifiers in the new space. In this paper, we consider the problem of learning image categorizers on large image sets (e.g. > 100k images) using bag-of-visual-words (BOV) image representations and Support Vector Machine classifiers. We experiment with three approaches to BOV embedding: 1) kernel PCA (kPCA), 2) a modified kPCA we propose for additive kernels and 3) random projections for shift-invariant kernels. We report experiments on 3 datasets: Cal-tech101, VOC07 and ImageNet. An important conclusion is that simply square-rooting BOV vectors - which corresponds to an exact mapping for the Bhattacharyya kernel - already leads to large improvements, often quite close to the best results obtained with additive kernels. Another conclusion is that, although it is possible to go beyond additive kernels, the embedding comes at a much higher cost.

A state-of-the-art approach to measure the similarity of two images is to model each image by a continuous distribution, generally a Gaussian mixture model (GMM), and to compute a probabilistic similarity between the GMMs. One limitation of traditional measures such as the Kullback-Leibler (KL) divergence and the probability product kernel (PPK) is that they measure a global match of distributions. This paper introduces a novel image representation. We propose to approximate an image, modeled by a GMM, as a convex combination of K reference image GMMs, and then to describe the image as the K-dimensional vector of mixture weights. The computed weights encode a similarity that favors local matches (i.e. matches of individual Gaussians) and is therefore fundamentally different from the KL or PPK. Although the computation of the mixture weights is a convex optimization problem, its direct optimization is difficult. We propose two approximate optimization algorithms: the first one based on traditional sampling methods, the second one based on a variational bound approximation of the true objective function. We apply this novel representation to the image categorization problem and compare its performance to traditional kernel-based methods. We demonstrate on the PASCAL VOC 2007 dataset a consistent increase in classification accuracy.

We introduce a novel family of contextual measures of similarity between distributions: the similarity between two distributions q and p is measured in the context of a third distribution u. In our framework any traditional measure of similarity / dissimilarity has its contextual counterpart. We show that for two important families of divergences (Bregman and Csisz'ar), the contextual similarity computation consists in solving a convex optimization problem. We focus on the case of multinomials and explain how to compute in practice the similarity for several well-known measures. These contextual measures are then applied to the image retrieval problem. In such a case, the context u is estimated from the neighbors of a query q. One of the main benefits of our approach lies in the fact that using different contexts, and especially contexts at multiple scales (i.e. broad and narrow contexts), provides different views on the same problem. Combining the different views can improve retrieval accuracy. We will show on two very different datasets (one of photographs, the other of document images) that the proposed measures have a relatively small positive impact on macro Average Precision (which measures purely ranking) and a large positive impact on micro Average Precision (which measures both ranking and consistency of the scores across multiple queries).

We present a novel approach to compute the similarity between two unordered variable-sized vector sets. To solve this problem, several authors have proposed to model each vector set with a Gaussian mixture model (GMM) and to compute a probabilistic measure of similarity between the GMMs. The main contribution of this paper is to model each vector set with a GMM adapted from a common ldquouniversalrdquo GMM using the maximum a posteriori (MAP) criterion. The advantages of this approach are twofold. MAP provides a more accurate estimate of the GMM parameters compared to standard maximum likelihood estimation (MLE) in the challenging case where the cardinality of the vector set is small. Moreover, there is a correspondence between the Gaussians of two GMMs adapted from a common distribution and one can take advantage of this fact to compute efficiently the probabilistic similarity. This work is applied to the image categorization problem: images are modeled as bags of low-level features and classification is performed using a kernel classifier based on the proposed similarity measure. Experimental results on the PASCAL VOC 2006 and VOC 2007 databases show the excellent performance of our approach.