This article details two approaches to compute barycenters of measures using 1-D Wasserstein distances along radial projections of the input measures. The first method makes use of the Radon ...transform of the measures, and the second is the solution of a convex optimization problem over the space of measures. We show several properties of these barycenters and explain their relationship. We show numerical approximation schemes based on a discrete Radon transform and on the resolution of a non-convex optimization problem. We explore the respective merits and drawbacks of each approach on applications to two image processing problems: color transfer and texture mixing.
This paper investigates a new stochastic algorithm to approximate semi-discrete optimal transport for large-scale problem, i.e., in high dimension and for a large number of points. The proposed ...technique relies on a hierarchical decomposition of the target discrete distribution and the transport map itself. A stochastic optimization algorithm is derived to estimate the parameters of the corresponding multi-layer weighted nearest neighbor model. This model allows for fast evaluation during synthesis and training, for which it exhibits faster empirical convergence. Several applications to patch-based image processing are investigated: texture synthesis, texture inpainting, and style transfer. The proposed models compare favorably to the state of the art, either in terms of image quality, computation time, or regarding the number of parameters. Additionally, they do not require any pixel-based optimization or training on a large dataset of natural images.
This paper proposes a new definition of the averaging of discrete probability distributions as a barycenter over the Monge-Kantorovich optimal transport space. To overcome the time complexity ...involved by the numerical solving of such problem, the original Wasserstein metric is replaced by a sliced approximation over 1D distributions. This enables us to introduce a new fast gradient descent algorithm to compute Wasserstein barycenters of point clouds.
This new notion of barycenter of probabilities is likely to find applications in computer vision where one wants to average features defined as distributions. We show an application to texture synthesis and mixing, where a texture is characterized by the distribution of the response to a multi-scale oriented filter bank. This leads to a simple way to navigate over a convex domain of color textures.
We propose GOTEX, a general framework for texture synthesis by optimization that constrains the statistical distribution of local features. While our model encompasses several existing texture ...models, we focus on the case where the comparison between feature distributions relies on optimal transport distances. We show that the semi-dual formulation of optimal transport allows to control the distribution of various possible features, even if these features live in a high-dimensional space. We then study the resulting minimax optimization problem, which corresponds to a Wasserstein generative model, for which the inner concave maximization problem can be solved with standard stochastic gradient methods. The alternate optimization algorithm is shown to be versatile in terms of applications, features and architecture; in particular, it allows to produce high-quality synthesized textures with different sets of features. We analyze the results obtained by constraining the distribution of patches or the distribution of responses to a pre-learned VGG neural network. We show that the patch representation can retrieve the desired textural aspect in a more precise manner. We also provide a detailed comparison with state-of-the-art texture synthesis methods. The GOTEX model based on patch features is also adapted to texture inpainting and texture interpolation. Finally, we show how to use our framework to learn a feed-forward neural network that can synthesize on-the-fly new textures of arbitrary size in a very fast manner. Experimental results and comparisons with the mainstream methods from the literature illustrate the relevance of the generative models learned with GOTEX.
State of the art deep generative networks have achieved such realism that they can be suspected of memorizing training images. It is why it is not uncommon to include visualizations of training set ...nearest neighbors, to suggest generated images are not simply memorized. We argue this is not sufficient and motivates studying overfitting of deep generators with more scrutiny. We address this question by i) showing how simple losses are highly effective at reconstructing images for deep generators ii) analyzing the statistics of reconstruction errors for training versus validation images. Using this methodology, we show that pure GAN models appear to generalize well, in contrast with those using hybrid adversarial losses, which are amongst the most widely applied generative methods. We also show that standard GAN evaluation metrics fail to capture memorization for some deep generators. Finally, we note the ramifications of memorization on data privacy. Considering the already widespread application of generative networks, we provide a step in the right direction towards the important yet incomplete picture of generative overfitting.
An Analysis of the SURF Method Oyallon, Edouard; Rabin, Julien
Image processing on line,
01/2015, Letnik:
5
Journal Article
Recenzirano
Odprti dostop
The SURF method (Speeded Up Robust Features) is a fast and robust algorithm for local, similarity invariant representation and comparison of images. Similarly to many other local descriptor-based ...approaches, interest points of a given image are defined as salient features from a scale-invariant representation. Such a multiple-scale analysis is provided by the convolution of the initial image with discrete kernels at several scales (box filters). The second step consists in building orientation invariant descriptors, by using local gradient statistics (intensity and orientation). The main interest of the SURF approach lies in its fast computation of operators using box filters, thus enabling real-time applications such as tracking and object recognition. The SURF framework described in this paper is based on the PhD thesis of H. Bay ETH Zurich, 2009, and more specifically on the paper co-written by H. Bay, A. Ess, T. Tuytelaars and L. Van Gool Computer Vision and Image Understanding, 110 (2008), pp. 346–359. An implementation is proposed and used to illustrate the approach for image matching. A short comparison with a state-of-the-art approach is also presented, the SIFT algorithm of D. Lowe International Journal of Computer Vision, 60 (2004), pp. 91–110, with which SURF shares a lot in common.
We investigate in this work a versatile convex framework for multiple image segmentation, relying on the regularized optimal mass transport theory. In this setting, several transport cost functions ...are considered and used to match statistical distributions of features. In practice, global multidimensional histograms are estimated from the segmented image regions and are compared to reference models that are either fixed histograms given a priori, or directly inferred in the non-supervised case. The different convex problems studied are solved efficiently using primal–dual algorithms. The proposed approach is generic and enables multiphase segmentation as well as co-segmentation of multiple images.
This contribution deals with the Heeger-Bergen pyramid-based texture analysis/synthesis algorithm. It brings a detailed explanation of the original algorithm tested on many characteristic examples. ...Our analysis reproduces the original results, but also brings a minor improvement concerning non-periodic textures. Inspired by visual perception theories, Heeger and Bergen proposed to characterize a texture by its first-order statistics of both its color and its responses to multiscale and multi-orientation filters, namely the steerable pyramid. The Heeger-Bergen algorithm consists in the following procedure: starting from a white noise image, histogram matchings are performed to the noise alternatively in both the image domain and steerable pyramid domain, so that the corresponding histograms match the ones of the input texture.
The recent advent of powerful generative models has triggered the renewed development of quantitative measures to assess the proximity of two probability distributions. As the scalar Fréchet ...inception distance remains popular, several methods have explored computing entire curves, which reveal the trade-off between the fidelity and variability of the first distribution with respect to the second one. Several of such variants have been proposed independently and while intuitively similar, their relationship has not yet been made explicit. In an effort to make the emerging picture of generative evaluation more clear, we propose a unification of four curves known respectively as: the precision-recall (PR) curve, the Lorenz curve, the receiver operating characteristic (ROC) curve and a special case of Rényi divergence frontiers. In addition, we discuss possible links between PR / Lorenz curves with the derivation of domain adaptation bounds.