Jornada CEA "Machine Learning y reconocimiento de patrones aplicado a Visión por Computador"
 
Lugar: Universidad Carlos III (Leganés, Madrid)
Fecha: 15 de Junio de 2009 (día previo a la la Jornada de Seguimiento de Proyectos (DPI-CICYT)

 

Tradicionalmente, los algoritmos de visión por computador han representado los objetos y el entorno a partir de modelos físicos o estadísticos con un alto componente de ajuste manual. Sin embargo, las nuevas tecnologías usadas para la adquisición de imágenes, junto con el gran aumento en la capacidad de cálculo de los ordenadores han permitido capturar y procesar grandes cantidades de información visual de forma eficiente. En este marco, se ha popularizado el uso de algoritmos de aprendizaje automático (machine learning en inglés) para, a partir de grandes bases de datos, construir modelos visuales mucho más robustos que los usados tradicionalmente. De esta forma, se han podido desarrollar sistemas basados en visión por computador altamente adaptables y abordar con éxito un sinfín de retos en áreas tan diversas como la categorización de objetos o la reconstrucción 3D de imágenes de neuronas.

Dentro del machine learning, se ha producido también un progreso notable en las técnicas de reconocimiento de patrones, tanto en el caso supervisado como en el no supervisado, así como, en los sistemas que combinan la visión o el análisis de imagen con otras técnicas inteligentes para clasificar o tomar decisiones.

En esta jornada de un solo día se presentarán los últimos avances en técnicas de machine learning y reconocimiento de patrones para visión por computador que se hayan desarrollado en universidades y grupos de investigación españoles. El principal objetivo es crear un foro de discusión y posibles colaboraciones entre los investigadores.

 

PROGRAMA:
 

Aula de grados de la Universidad Carlos III de Madrid (http://jornrobovis2009.uc3m.es/localizacion)

 
ACTUALIZADO: Añadidas las presentaciones de las ponencias
 

Hora

Ponencia

Presentación

12:00 - 12:05

Presentación de la Jornada

 

12:05 - 12:25

Angela Ribeiro y Xavier P. Burgos-Artizzu
"Algoritmos Genéticos en Visión por Computador. Aplicaciones". Instituto de Automática Industrial/Consejo Superior de Investigaciones Científicas (IAI-CSIC)

 

12:25 - 12:45

Germán González
“Learning Features for Filament Detection”.
EPFL, (École Polytéchnique Fédérale de Lausanne, Suiza)

Descarga PDF

12:45 - 13:05

Víctor González Castro
“Cuantificación de la proporción de células de esperma dañadas utilizando análisis de imagen y redes neuronales”.
Universidad de León.

Descarga PDF

13:05 - 13:25

Enrique Cabello Pardos
“Sistema de visión de detección de manos en conductores de vehículos”
Universidad Rey Juan Carlos.

 

13:25 - 15:25

Comida

 

15:25 - 15:45

Enrique Coiras
“Underwater Imaging Systems for Maritime Security” .
NATO Undersea Research Centre (NURC, http://www.nurc.nato.int)

Descarga PDF

15:45 - 16:00

Café

 

16:00 - 16:20

Jose J. Guerrero
“Visual door detection integrating appearance and shape cues”.
Universidad de Zaragoza

Descarga PDF

16:20 - 16:40

Óscar Chang
“Evolving Cooperative Agents for Controlling a Vision Guided Mobile Robots”.
ETSII - Universidad Politécnica de Madrid

 

16:40 - 17:00

David Martín
“Fusión de imágenes láser y cámara para caracterización de obstáculos en exteriores”.
Instituto de Automática Industrial/Consejo Superior de Investigaciones Científicas (IAI-CSIC)

 

17:00 - 17:20

X. Boix, J. Serrat
“Prior learning for dense disparity from stereo pair sequences”.

Descarga PDF

17:20 - 17:50

Discusión final

 

 

Organizadores

José María Martínez Montiel
I3A - Universidad de Zaragoza

josemari@unizar.es

Enrique Alegre Gutiérrez
Universidad de León

enrique.alegre@unileon.es

Francesc Moreno-Noguer
Institut de Robòtica i Informàtica Industrial, CSIC-UPC

fmoreno@iri.upc.edu

 

Germán González: “Learning Features for Filament Detection”. EPFL, École Polytechnique Fédérale de Lausanne, Suiza)

State-of-the-art approaches for detecting filament-like structures in noisy images rely on filters optimized for signals of a particular shape, such as an ideal edge or ridge. While these approaches are optimal when the image conforms to these ideal shapes, their performance quickly degrades on many types of real data where the image deviates from the ideal model, and when noise processes violate a Gaussian assumption.
We show that by using a learning approach, we can outperform state-of-the-art filament detection techniques on many different kinds of imagery. More specifically, we demonstrate superior performance for the detection of blood vessel in retinal scans, neurons in brightfield microscopy imagery, and streets in satellite imagery.

 

David Martín: “Fusión de imágenes láser y cámara para caracterización de obstáculos en exteriores”. Instituto de Automática Industrial/Consejo Superior de Investigaciones Científicas (IAI-CSIC)

El objetivo del trabajo es el desarrollo de algoritmos para combinar catracteristicas de objetos extraidas de un láser de barrido 2D y de una cámara de visión. Se presenta la segmentación de las imágenes, calibración y extracción de características para hacer corresponder caracterñisticas extraidas de la imagen visual con perfiles de profundidad proporcionados por los barridos del láser.
Estas características son de gran utilidad en el pilotaje reactivo de un tractor comercial adaptado para la realización semi-autonoma de tareas en campo. El conocimiento extraido de objetos imprevistos en la trajectoria del tractor hacia un objetivo, se utilizan para la activación de los agentes {AVANZAR, EVITAR, PARAR} de la capa reactiva de la arquitectura de control basada en agentes de comportamiento.

 

Víctor González Castro: Cuantificación de la proporción de células de esperma dañadas utilizando análisis de imagen y redes neuronales. (Quantifying the proportion of damaged sperm cells based on image analysis and Neural Networks)”. Universidad de León

Insemination techniques in the veterinary field demand more objective methods to control the quality of sperm samples. In particular, different factors may damage a number of sperm cells that is difficult to predict in advance. This paper addresses the problem of quantifying the proportion of damaged/intact sperm cells in a given sample based on computer vision techniques and supervised learning. Unlike common supervised classification approaches, neither the individual example classification is important nor the class distribution assumed in learning can be considered stationary. To deal with this, an estimation process based on Posterior Probability estimates (PP), and known to increase the classifier accuracy under changes in class distributions, is assessed here for quantification purposes. It is compared with two approaches based on the classifier confusion matrix (Adjusted Count andMedian Sweep) and the naive approach based on classifying and counting. Experimental results with boar sperm samples and back propagation neural networks show that the PP based quantification outperforms any of the previously considered approaches in terms of the Mean Absolute Error, Kullback Leibler divergence and Mean Relative Error. Moreover, in spite of an imperfect classification, the PP approach guarantees a uniform Mean Absolute Error (around 3%) for whatever combination of training and test class distributions, what is very promising in this practical field.

 

Óscar Chang: “Evolving Cooperative Agents for Controlling a Vision Guided Mobile Robots”. ETSII - Universidad Politécnica de Madrid

We have studied and developed the behavior of two specific neural processes, for vehicle driving and path planning, used to control mobile robots. Each processor behaves as an independent agent and is an instance of a neural network trained for a definite task. Through simulated evolution full grown agents are encouraged to socialize by opening low bandwidth, asynchronous channels between them. The results indicate that under evolutive pressure controlling agents can spontaneously develop agent-communication skills (protolanguage) that make sense of the agents´ interchanged information. The emerged cooperative behavior raises the level of competence of vision guided mobile robots and allows a convenient exploration of the environment. The system has been tested in a simulated environment and shows a robust performance.

 

Enrique Coiras: "Underwater Imaging Systems for Maritime Security". NATO Undersea Research Centre (NURC, http://www.nurc.nato.int)

The observation of the submarine environment is complicated by several important limitations. Since the transmission of electromagnetic waves is drastically reduced underwater, visual cameras, GPS and radio transmissions are completely ineffective. The only means of imaging and communication underwater is to use sound. Sonar (SOund NAvigation and Ranging) systems have greatly evolved since their inception in the 1950s and are now able to offer almost photographic quality images of the seafloor. This permits to apply traditional image processing techniques to detect and classify objects of interest, such as underwater mines. Other applications, like shape-from-shading to obtain 3D information, offer new possibilities for inspection and analysis of archaeological sites and critical infrastructures.
In this presentation, the underwater observation systems used by NURC are introduced and their main applications discussed, making special emphasis on the new Synthetic Aperture Sonar of the MUSCLE autonomous vehicle, capable of achieving imaging resolutions of 2.5cm x 2.5cm at up to 200m range.

 

Jose J. Guerrero: “Visual door detection integrating appearance and shape cues”. Universidad de Zaragoza.

An important component of human–robot interaction is the capability to associate semantic concepts with encountered locations and objects. This functionality is essential for visually guided navigation as well as location and object recognition. In this paper we focus on the problem of door detection using visual information only. Doors are frequently encountered in structured man-made environments and function as transitions between different places. We adopt a probabilistic approach for door detection, by defining the likelihood of various features for generated door hypotheses. Differing from previous approaches, the proposed model captures both the shape and appearance of the door. This is learned from a few training examples, exploiting additional assumptions about the structure of indoor environments. After the learning stage, we describe a hypothesis generation process and several approaches to evaluate the likelihood of the generated hypotheses. The approach is tested on numerous examples of indoor environment. It shows a good performance provided that the door extent in the images is sufficiently large and well supported by low level feature measurements.

 

X. Boix, J. Serrat: “Prior learning for dense disparity from stereo pair sequences”.

Our work addresses the problem of dense disparity estimation from a pair of stereo sequences through maximum a posteriori inference on a Markov random field (MRF). We study the modelling and learning of the priors involved in the MRF design. Three different state-of-the-art learning methods are compared through a new proposed order-two model.
If we state the learning as a log-likelihood maximization, the estimation of the partition function has to be computed. We compare two methods that deal with this approximation, Sharstein and Pal [1], and the contrastive divergence of Hinton [2]. In contrast, Samuel and Tappen [3] state the parameter learning so that the minimum energy solution is as similar as possible to the ground-truth. The experimental results illustrate the potential of learning automatically the parameters of models with richer structure than standard hand-tuned MRF models..

 

 

 

Logo CEA-IFAC Logo MICINN Logo UC3M