Monthly Archives: May 2010

Yet another report by Olivier who attended RFIA 2010

RFIA 2010 — Reconnaissance de formes et intelligence artificielle

RFIA (which stands for Reconnaissance de formes et intelligence artificielle) is a French-speaking conference about pattern recognition and artificial intelligence. The 17th edition held in Caen from January 20 to January 22 2010.

There was four parts in the workshop: an invited speaker session, a short talk session, a poster session and a demo session.

The short talk session was divided into two tracks: pattern recognition and artificial intelligence. I mainly followed the Pattern Recognition track. Unfortunately I was only able to attend the first two days of the conference, so I will present only talks from these days. Sang Ly presented a method to estimate moves of a stereoscopic vision system. Hervé Jégou presented a new compact representation of bag-of-words for image retrieval. Thierry Germa presented some results to follow people with a mobile platform, using information from image and RFID tags. Pierre Lébraly explained an extrinsic calibrating method for multi-camera system, using a planar mirror.

The three invited talks I attended were given by Olivier Teytaud (INRIA Saclay, France) speaking about artificial intelligence methods for Go game, Schlomo Zilberstein (University of Massachusetts, USA) presenting challenges and directions for decentralized decision making and by Nicu Sebe (University of Amsterdam, Nederlands) detailing some perspectives for human centered computing.

Frank.

Olivier attended the VISAPP conference where he presented results on

Reranking with Contextual dissimilarity measures from representational Bregman k-means

Here is the abstract followed by a short report of the conference.

We present a novel reranking framework for Content Based Image Retrieval (CBIR) systems based on con-textual dissimilarity measures. Our work revisit and extend the method of Perronnin et al. (Perronnin et al., 2009) which introduces a way to build contexts used in turn to design contextual dissimilarity measures for reranking. Instead of using truncated rank lists from a CBIR engine as contexts, we rather use a clustering algorithm to group similar images from the rank list. We introduce the representational Bregman divergences and further generalize the Bregman k-means clustering by considering an embedding representation. These representation functions allows one to interpret ?-divergences/projections as Bregman divergences/projections on ?-representations. Finally, we validate our approach by presenting some experimental results on ranking performances on the INRIA Holidays database.

VISAPP 2010 — International Conference on Computer Vision Theory and Applications

VISAPP is part of VISIGRAPP – The International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. The purpose of this joint conference is to bring together researchers and practitioners on the areas of computer vision, imaging, computer graphics and information visualization, interested in both theoretical advances and applications in these fields. Computer Vision, Imaging, Computer Graphics and Information Visualization are well known areas which are becoming more and more interrelated with important interdisciplinary work, often as a result of an iterative combined process of image analysis and synthesis with models created in one of the fields being used to improve models created in another.

The VISIGRAPP component conferences are specialized in the following topics: GRAPP is structured along four main tracks, covering different aspects related to Computer Graphics, from Modelling to Rendering, including Animation and Interactive Environments, IMAGAPP covers theory, applications and technologies related to image display, colour coding, medical imaging, remote sensing, business document processing, digital fabrication, printing and electronic devices, VISAPP has also four main tracks, namely: Image Formation and Processing, Image Analysis, Image Understanding and Motion, Tracking and Stereo Vision and IVAPP structured along several topics related to Information Visualization.

The 2010 edition of VISIGRAPP was held in Angers (France) from May 17th to May 21th. There were 4 invited speakers: Pascal Fua, from the École Polytechnique Fédérale de Lausanne (Switzerland), spoke about different ways of modeling deformable surfaces from monocular video sequences; Ali Mohammad-Djafari, from CNRS and Supéléc (France) spoke about methods for solving inverse problems, mainly regularization and bayesian estimation approaches, with applications in imaging and computer vision. The talk by Gabriela Csurka, from Xerox Research (France), was about Fisher kernel representation, describing Bag-Of-Features representation, Fisher vector and some applications. The last keynote speaker, Brian Barsky from University of California Berkeley, described recent research for modeling the human vision process, mainly for simulating eye diseases and consequences of surgery cures.

There were a lot of parallel sessions (2 or 3 only for VISAPP) which led to difficult choices between interesting talks. I will just a say a few words about some talks from the session I attended: Nuria Ortigosa presented the paper “Disparity maps for free path detection” describing an algorithm to detect obstacles and free path using stereo images. Sang Min presented the paper “Object retrieval based on user-drawn sketches”, describing a similarity measure between sketches based on tensorial fields. Duan-Yu Chen, presenting the paper “Real-time gender recognition for uncontrolled environment of real-life images” spoke about a way to combine different approaches in order to get a robust algorithm. Laura Papaelo presented TOPMESH, a powerful library to build models of non-manifolds objects. Codruta Orniana-Ancuti, presenting “Robust grayscale conversion for vision-substitution systems”, described a method to convert color images to grayscale images suitable for an auditory-vision system. Roland Moerzinger presented the paper “Improving person detection in videos by automatic scene adaptation” which introduces a way to improve person detection quality in the case of a static camera and a single planar ground. Zhaolin Su, for the paper “Real-time enhancement of image and video saliency using semantic depth of field”, presented a way to change the depth of field of an image using saliency information. The talk of Svenja Kahn, “Time-of-flight based scene reconstruction with a mesh processing tool for model based camera tracking”, described a way to build 3D models using a time-of-flight camera without the need of a prior knowledge of the model, for application in augmented reality.

Even if the goal to allow meetings between people from different fields is very interesting, it was really difficult to attend sessions from other fields without missing interesting talks in one’s own field.

The Jensen-Bregman divergence is a distortion measure defined by the Jensen difference provided by a strictly convex function. Jensen-Bregman divergences extend the well-known Jensen-Shannon divergence by allowing to choose an arbitrary convex function generator instead of the standard Shannon entropy. This class of Jensen-Bregman divergences notably includes the squared Euclidean distance. Although Jensen-Bregman divergences are symmetric distances by construction, they are not metrics as they violate the triangle inequality. We study the geometric properties and combinatorial complexities of both the Voronoi diagrams and the centroidal Voronoi diagrams induced by such as class of information-theoretic divergences. We show that those Jensen- Bregman divergences appear naturally into two contexts: (1) when symmetrizing Bregman divergences, and (2) when computing the Bhattacharyya distances of statistical distributions. The Bhattacharyya distance measures the amount of overlap of distributions, and is popularly used to provide both lower and upper bounds in machine learning: the Bayes? misclassification error. Since the Bhattacharyya distance of popular distributions in statistics called the exponential families (including familiar Gaussian, Poisson, multinomial, Beta/Gamma families, etc.) can be computed equivalently as Jensen-Bregman divergences, the Jensen-Bregman Voronoi diagrams allow one also to study statistical Voronoi diagrams induced by an entropic function.