April, 6th

April, 7th

April, 8th

14:00h-14:30h

Sylvie Philipp-Foliguet (ENSEA)

"FReBIR:  Fuzzy Region-Based Image Retrieval"

Jean Cousty (ESIEE)

"On Watershed Cuts" 

André Saúde (UFLA)

"Watershed and Image Restoration"

14:30h-15:00h

Eduardo Valle (UFMG)

"Indexing High-Dimensional Data - Application to CBIR"

15:00h-15:30h

Marcelo Bernardes Vieira (UFJF)

 "Multiresolution Decomposition and Visualization of
3D Scalar and Tensor Fields"  

Marcos André Gonçalves (UFMG)

 "Learning to Rank at Query-Time Using Association Rules" 

Júlia Epischina Engrácia Oliveira (UFMG)
"Content-based Image Retrieval of Mammographies using the IRMA database"

16:00h-16:30h

Ana Paula Brandão Lopes (UFMG)
"Human Actions Recognition"

16:30h-17:00h

 Sandra Eliza Fontes de Avila (UFMG)
"VSUMM: An Approach Based on Color Feature Extraction for Automatic Summarization and Subjective Evaluation of Static Video Summaries"

17:00h-17:30h

FReBIR: Fuzzy Region-Based Image Retrieval

Presenter:

Sylvie Philipp-Foliguet

FReBIR is a method of image indexing and retrieval which takes into account the relative positions of the regions within the image. Indexing is based on a fuzzy segmentation of the image. Fuzzy regions are then indexed by colour and texture. The image retrieval is based on inexact graph matching, taking into account both the similarity between regions and the spatial relation between them. We propose, on one hand a solution to reduce the combinatorial complexity of the graph matching, and on the other hand, several measures of similarity between graphs allowing the result images ranking. Similarity measures use kernel functions adapted to vectors, bags of features or graphs. SVM classifiers are used through  relevance feedback loops to retrieve categories of images.

Applications concern image and 3D object retrieval. The method is adapted to partial queries, aiming for example at retrieving images containing a specific type of object.
[top]

Multiresolution Decomposition and Visualization of 3D Scalar and Tensor Fields

Presenter:

Marcelo Bernardes Vieira

The visualization of 3D tensor and scalar fields is a challenging task. Firstly, the volumetric data has inner details that should not be occluded by outter parts or object surfaces. The second problem is that tensors capture multivariate data that should be expressed accordingly. Naive approaches for showing them often result in useless image information. Multiresolution decomposition is a well known method for signal analysis and has been extensively used in computer graphics problems. In a recently opened research line, we investigate how multiresolution can be applied to assess and view multivariate data. With promising results, we present a method for viewing multiresolution edges in images and a method to investigate the multiresolution structure of general 3D tensor fields.
[top]

VSUMM: An Approach Based on Color Feature Extraction for Automatic Summarization and Subjective Evaluation of Static Video Summaries

Presenter:

Sandra Eliza Fontes de Avila

Abstract:

Advances in compression techniques, in decreasing cost of storage, and in high-speed transmission have facilitated the way videos are created, stored and distributed. The increase in the amount of video data deployed and used in many applications, such as search engines and digital libraries, reveals not only the importance as multimedia data type, but also leads to the requirement of efficient management of video data. This management paved the way for new research areas, such as automatic video summarization. Essentially, this research area consists of automatic generating a short summary of a video, which can either be a static summary (keyframes set) or a dynamic summary (set of video segments). This work presents VSUMM, a methodology for the development of static summaries. The method is based on the extraction of color-based features from video frames and unsupervised classification. The video summaries produced are evaluated by users and compared with approaches found in the literature. With a confidence level of 98%, the proposed solution provided results with superior quality relative to the approaches to which it was compared.
[top]

Human Actions Recognition

Presenter:

Ana Paula Brandão Lopes

Abstract:

The ability to automatically recognize human actions directly from video information has many potential applications, like improving video content-based indexing and retrieval, identifying suspect behavior in surveillance scenarios, remotely monitoring elderly people or analyzing sports videos, for example. In this presentation, we provide an overview of the different representation approaches for human action recognition. Then, we provide some detail for a promising one, namely, bag of visual features (BOVF). Finally, we show some results we achieved with a BOVF implementation of ours in a standard human actions database.

[top]

Bag of Visual Features Applications

Presenter:
Ana Paula Brandão Lopes
 
Abstract:

Bag of visual features (BOVF) representations have been used successfully in several tasks, like object and human actions recognition and scene classification. In this presentation, we review the basics of BOVF and show how we are applying BOVF to two specific applications: classification of historical photographs based on the presence of buildings and nude detection.

[top]

On Watershed Cuts

Presenter:
Jean Cousty
 
Abstract:

The watershed transform is an efficient and popular tool for image segmentation. In this talk, we study the watersheds in edge-weighted graphs. We define the watershed cuts following the intuitive idea of drops of water flowing on a topographic surface.

In a first part, we establish the consistency of these watersheds: they can be equivalently defined by their catchment basins (through a steepest descent property) or by the dividing lines separating these catchment basins (through the drop of water principle). Then, we prove, through an equivalence theorem, their optimality in terms of minimum spanning forests.

In a second part, we present a thinning paradigm from which we derive three algorithmic watershed cut strategies: the first one is well suited to parallel implementations, the second one leads to a flexible linear-time sequential implementation whereas the third one links the watershed cuts and the popular flooding algorithms.

In the third part of the talk, we state that the watershed cuts preserve a notion of contrast, called connection value, on which are (implicitly) based several morphological region merging methods. This leads us to establish the links and differences between watershed cuts, minimum spanning forests and shortest-path forests.

Finally, we conclude the talk by showing illustrations of the proposed framework to the segmentation of grayscale images, artwork surfaces and diffusion tensor images.

[top]

Learning to Rank at Query-Time Using Association Rules

Presenter:
Marcos André Gonçalves
 
Abstract:

Some applications have to present their results in the form of ranked lists. This is the case of many information retrieval applications, including Content-Based Image Retrieval (CBIR), in which objects (documents, images) must be sorted according to their relevance to a given query. This has led the interest of the  information retrieval community in  methods that automatically learn effective ranking functions. In this paper we propose a novel method which uncovers patterns (or rules) in the training data associating features of the object with its relevance to the query, and then uses the discovered rules to rank these objects. To address typical problems that are inherent to the utilization of association rules (such as missing rules and rule explosion), the proposed method generates rules on a demand-driven basis, at query-time. The result is an extremely fast and effective ranking method. We conducted a systematic evaluation of the proposed method using the LETOR benchmark collections. We show that generating rules on a demand-driven basis can boost ranking performance, providing gains ranging from 12% to 123%, outperforming the state-of-the-art methods that learn to rank, with no need of time-consuming and laborious pre-processing. As a highlight, we also show that additional information, such as query terms, can make the generated rules more discriminative, further improving ranking performance. Despite focused in document retrieval, the techniques presented here are easily adapted to any retrieval task in which the objects are represented as a bag of features.

[top]

OpenMoCap: An Open Source Software for Optical Motion Capture

Presenter:
David Lunardi Flam and João Victor Boechat Gomide

Abstract:
This work presents the actual development stage of an open source software for motion capture, based on computer image analysis techniques. The code is being written in C++, with OpenCV and QT libraries, aiming real time applications. Currently, we have almost completed the full pipeline for motion capture with markers and now working on the final step, which is direct 3d metric reconstruction. We are testing the architecture using Natural Point’s V100 cameras with infrared LEDs and retroreflective balls. Future steps include comparison of our system with commercial tracking tools and output moCap data in BVH format. Also, we are modeling a short film that will be animated using OpenMoCap.
[top]

Event Detection in Surveillance Videos

Presenter:
Guillermo Cámara Chávez

Abstract:
Large quantities of video surveillance data exist in today’s world. Cameras are everywhere constantly recording daily occurrences from many angles. Our objective is to develop a framework to aid video analysts in detecting suspicious activity. We consider an activity based video content representation. Visual events are detected and classified automatically in the scene. Although there are many ways to represent the content of video clips in current video retrieval algorithms, there still exists a semantic gap between users and retrieval systems. Visual surveillance systems supply a platform for investigating semantic-based video retrieval.
[top]

Watershed and Image Restoration

Presenter:

André Saúde

Abstract:

Image restoration has been frequently treated as a global optimization problem. Usually, the restored image is a function that minimizes the global energy that models the noise. One solution available in the literature is to reduce the energy minimization problem to the problem of finding a minimal cut in a graph, which is a classical problem. After a fine analysis of the minimal cut solution, we propose some greedy assumptions in order to treat the problem as a local optimization problem, and in consequence to obtain a fast algorithm that computes an aproximation of the global minimum. With such greedy assumptions, the image restoration could be computed by watershed-like algorithms. This work is under research. The aim is to share this experience with the colleagues as a way to get feedback about the solution we propose.

[top]

Indexing High-Dimensional Data - Application to CBIR

Presenter:

Eduardo Valle

Abstract:

The troubles of multimedia information retrieval start at its most elementary operation : matching the high-dimensional feature vectors used to describe the data. In this talk, we will discuss how recent innovative methods are taming the infamous "curse of dimensionality" and how they can be used in CBIR. The author will discuss his recent contributions to the advance of the state-of-art and his current research endeavours.

[top]

Content-based Image Retrieval of Mammographies Using the IRMA Database

Presenter:

Abstract:
This work aims at developing a content-based image retrieval (CBIR) system for mammographies using as pattern the breast density. The database used, from the IRMA project, provides images with the ground truth already set, in a way to facilitate the evaluation of the proposed system. This work focuses on the breast density characterization through the texture representation together with the reduction of the dimensionality of the feature vector. Two methods are proposed: MammoSVD (using the singular value decomposition - SVD) and MammoPCA (using the two dimensional principal component analysis - 2DPCA). Support vector machine (SVM) is used for the retrieval task.
[top]

Automatic Detection of the Damaged Leaf Area by Pests in Cultivars Through Digital Images

 Presenter:

Antônio Carlos de Nazaré Júnior

Abstract:
In agriculture, there are many difficulties involved in the handling of pests in the tillage. The are several types of attack of plagues. One of them affects directly the leaf of the plant. The preservation of leaf is of much importance for the primary metabolism of the plant and for the maintenance of the production of fruits. Therefore, the accurate and precise detection of the damaged leaf area is essential for the determination of the control action, once a small damage does not require any control measures.

In the present work, we propose a method for automatic detection and/or quantification of the damaged leaf area by pests. These methods work with digital images captured through a digital camera (by farmers themselves) and/or scanners. Thus, they can overcome the difficulties presented by other methods in the literature, such as the use of the planimeter (a gridded surface used to measure the area of an arbitrary two-dimensional shape) and the recover of concave regions presented in the damaged leaves.

These new methods will reduce the time required for the assessment of the damage, and increase the reliability of the leaf analysis. At present, we are testing the methods in samples of soybean leaves (Glycine Max (L.) Merril) collected from the experimental fields of the Phytotechnical Department of the Federal University of Viçosa.

We are currently studying  pre-processing techniques for the suppression of noise and shadow (added to the images during the digitization of the samples), and also methods capable of detecting broken points in the edges of the leaves (necessary for a perfect quantification of the damaged leaf area).

[top]

An Approach for Photometric Validation in On-board Systems

Presenter:

Alexandre Wagner

Abstract:

Visual Information still represents one of the most common ways of interaction between a machine and a human being. In order to facilitate this interaction, machines are equipped with luminous components that form an on-board environment. In vehicles, interaction with conductors is made through the reading of the cluster information, the radio operability, and any other luminous component. Hence, it is very important that the internal lighting of a vehicle is in good harmony with the customer. To achieve this harmony, in this work photometric characteristics of the components, such as intensity, color, and homogeneity, are studied and measured. The goal of this work is to develop a methodology, based on the human visual perception, to automatically identify and quantify non-homogeneous regions of the lighting distribution in on-board systems. This is going to be done through the analysis of lighting components in digital images.

[top]

A Semi Automatic Methodology for Segmentation of the Coronary Artery Tree from Angiography

Presenter:

Daniel da Silva Diogo Lara

Abstract:

Nowadays, medical diagnostics using images has a considerable importance in many areas of medicine. It promotes and makes easier the ways for acquisition, transmission and analysis of medical images. The use of digital images in the medical area is still growing up and new application modalities are always appearing. Coronary Artery Disease (CAD) is the narrowing or blockage of arteries that provide the heart muscle with blood. Coronary angiography remains to be an indispensable tool in clinics today for the diagnosis of CAD. A fundamental component of a semi automatic angiography analysis is vessel detection. Vessel detection is a recognition problem that is challenging due to the complex nature of vascular trees and to imaging imperfections. One inherent imperfection of angiography is the intensity inhomogeneity between the larger and smaller vessels. Another imperfection common among many angiographic methods is the leakage of contrast agent into the background tissue that reduces the contrast between vessels and tissue. This work presents a developing methodology for a semi automatic segmentation of the coronary artery tree from angiography.

[top]