Google image-swirl and some ideas

I always asked to myself why Google and other search engines do not offered to the user new interaction/visualization mechanisms whether there have been many initiatives in this field. Actually, one of the first papers appeared almost ten years ago [Chen00]. Frofesor Fabio said to me that new technologies take much time to appear due to they are subject to "the convergence". Well, he had reason ;). To my surprise, Google recently published a new demo lab tool, called image-swirl, that allow the user to visually interact with its large image repository. In this screen shot, I show an example of the system. This initiative is very nice and motivating for my research.

I was searching by "carcinoma". The interaction begins with a set of piles of images organized according to cluster of similar images. One you select one pile, it goes to the front (expanded in a circle) whilst the rest go to the background. While you interact with the collection, a spiral shape preserves the navigation path. This tool has what an Information Visualization (IV) metaphor must have, which Ben Shneiderman calls IV mantra: “Overview, zoom and filter, details on demand” [1].

A visualization metaphor depends on the domain context and on the task. I am not sure that it is at all suitable to medical domain. That is, imagine a pathology's student navigating a collection of basal cell carcinoma images in order to learn about the presence of a pathology concept. He/her probably will need a visualization/interaction tool that allow him/her to do this task easily and quickly. I think that is a good research question what is a good visualization metaphor to this case, and how it can improve learning tasks. We could address this question in a master thesis :).

It is difficult to build a good similarity function to this concept using only low-level features (color, textures, borders) or combination of them with a priori information provided by the expert. One possibility to reduce this semantic gap is trying to fusion different information sources (text, images, audio, etc.) to construct this similarity function (what Juanca is doing), and another possibility is to reduce this semantic gap but learning from the user interaction. That is, we begin with a basic similarity function and through the user interaction we adapt this similarity function.

I have been reviewing the literature from information retrieval perspective about learning the similarity function through user interaction. I found that this area has been very active since some years ago. Actually, I found some approaches where the kernel matrix changes in each iteration, and I also found some papers addressing this issue but in the medical context.

I will talk about my reviewing in more detail in the next post.

Enjoy image-swirl.

;)

[1] Ben Shneiderman. The Eyes Have It: A Task by Data Type Taxonomy for Information Visualizations. In Proc. 1996 IEEE Symposium on Visual Languages (VL’96), pages 336–343. IEEE Computer Society, Boulder, Colorado. doi:10.1109/VL.1996.545307. http://citeseer.ist.psu.edu/shneiderman96eyes.html.

BiMed poster for SIPAIM

This is the poster about BiMed we will present at SIPAIM, although it is already printed, suggestions are welcome ;).

Download PDF

CIARP 2009

Yesterday, I arrived from Mexico. I attended to the CIARP conference to present the paper A Multi-class Kernel Alignment Method for Image Collection Summarization. The conference was really interesting, there were good works, contacts and tutorials.

I want to highlight in this blog four main things: feedback, related work, contacts and some photos :|.

Feedback
The oral presentation was fine ;). The time allocated was sufficient to convey the main ideas of the article. In the question's part I answered them satisfactory. I feel the work was interesting for the audience. Actually, the Director of CENATAV (Centro de Aplicaciones de Tecnologías de Avanzada) approached me at the end of my presentation, to invite Colombia belongs to the International Association for Pattern Recognition (IACR). I will describe it with more details in contact's section.

Related work
In the conference there were too many works related to image processing, some about robot vision, analysis of signal, video segmentation and tracking, computer vision and document processing. Here the conference proceedings.

With respect to my research, the following ones are the most relevant:

• A Brief Index for Proximity Searching. Eric Sadit Téllez et al.
  In this work, authors propose a method to quickly find similar objects in large databases. They propose to create an index structure and an algorithm based on permutations. Basically, they performs approximate search in metric spaces using the k nearest neighbor strategy.
• Automatic Choice of the Number of Nearest Neighbors in Locally Linear Embedding. Juliana Valencia-Aguirre et al.
  In this paper, authors propose to automatically find the number of nearest neighbors k in the computation of the LLE (Locally Linear Embedding) algorithm, a method for dimensionality reduction. The define a cost function that quantifies the quality of the embedding results. Normally, the k value is set manually in the algorithm which may cause that the results are not the best. They visually show how can be found this parameter and experiment with artificial and real-world data sets.
• Learning an Efficient Texture Model by Supervised Nonlinear Dimensionality Reduction Methods. Elnaz Barshan et al.
  Authors investigate the problem of texture recognition under varying lighting and viewing conditions. They use textons [1], which describe local properties of textures. The high dimensionality of the textons and feature histograms make it necessary to find suitable methods in order to address the dimensionality reduction tasks, that is, methods that allow to guide algorithms making use of a priori information (unsupervised learning). Authors use a supervised nonlinear dimensionality reduction method called CMVU (Colored Maximum Variance Unfolding) [2].

Other interesting works:

• Randomized Probabilistic Latent Semantic Analysis for Scene Recognition. Erik Rodner and Joachim Denzler.
• Dynamic Image Segmentation Method Using Hierarchical Clustering. Jorge Galbiati et al.
• A New Unsupervised Learning for Clustering Using Geometric Associative Memories. Benjamín Cruz et al.
• Simple Noise Robust Feature Vector Selection Method for Speaker Recognition. Gabriel Hernández et al.
• A Speed-Up Hierarchical Compact Clustering Algorithm for Dynamic Document Collections. Reynaldo Gil-García et al.
• A New Incremental Algorithm for Overlapped Clustering. Ariel Pérez Suárez.
• Improved Online Support Vector Machines Spam Filtering Using String Kernels. Ola Amayri and Nizar Bouguila.
  

Contacts

• José Ruiz Shulcloper. He is the Director of the CENETAV and member of the IAPR. He propose to me to involve Colombia as member of the IAPR, that is, to create the Colombian Society of Pattern Recognition. There are various countries in this association and he said to me that Colombia should be also member. The idea is to talk with professor Fabio about our interest in belonging to this association. If we are interested on, we have to write to José Ruiz and he will send to us what are the requirements.
• Hector Allende. He is president of the Chilean Society of Pattern Recognition and Director of the Doctorate Program at Universidad Técnica Federico Santa María (UTFSM). He works in Machine Learning and Computational Intelligence. He said to us that he works in base ML research and supports other groups in applied research. I think that he is an importa
  
• Gabriel Hernández. Gabriel is a PhD student of the CENETAV. He is working on Pattern Recognition and presented an interesting work in the conference. His work is focused on Speaker Recognition.
  
• Carlos Martínez. Carlos is a PhD student at University of Exeter (UK). He works on the field of 2D and 3D image processing. The topic of his thesis is based on developing shape descriptors both for 2D and 3D shapes, and test their possible applications for image analysis, classification, and recognition.

Photos

Well, after an interesting conference it was necessary some of tourism. Here some photos of Teotihuacan:

Pyramid of the Sun
























The Pyramid of the Sun and the Avenue of the Dead from the Pyramid of the moon.












References

[1] Leung, T.K., Malik, J. Representing and recognizing the visual appearance of materials using three-dimensional textons. International Journal of Computer Vision 43(1), 29-44 (2001).

[2] Smola, A.J., Borgwardt, K.M., Song, L., Gretton, A. Colored maximum variance unfolding. In: NIPS (2007).

Microsoft eScience presentation

The Workshop has an interesting program: Carnigie Mellon, Oxford, Harvard, ... , Universidad Nacional de Colombia

http://research.microsoft.com/en-us/events/escience2009/agenda.aspx

Universidad Nacional de Colombia: "Presente, presente, presente."

MedViz and BiMed videos

I published two videos were MedViz and BiMed functionalities are shown. They are online at http://camargoj.googlepages.com (Software section).

Server migration

The server migration was not easy. It was necessary to collect the source code of all modules, change parameters, compile and redeploy. But finally, it is on-line. http://www.informed.unal.edu.co

New version of MedViz

I published a new version of the MedViz system.

MedViz is a web tool for visualizing and exploring large collection of medical images. In this prototype, I provide interaction functionalities that allow the user to navigate the collection.

Functionality
* In the screen are displayed 100 images of the Corel data set (2500 images)
* Select one image as relevant (dblclick)
* A new summary is generated with the "Start again" button
* Retrieve images pressing the "Search" button
* The system will retrieve new images according to the images selected as relevants by the user

Interacción2009

This week I am attending to the Interacción2009 conference. My presentation was nice. In general, the conference topics are related to HCI theory, usability, collaborative work and awareness. My talk was a little different because it had machine learning and maths equations, I think that it was interesting to the people. Cesar Collazos, the conference chair, suggested to me the following ideas:

• Collaborative image collection annotation to improve the image collectio visualization
• New interfaces to visualize the image collection, for example, mobile devices such as iPhone, etc.

Well, tomorrow more conference news.

Informed meeting

Today we were working with informed development team. We defined some activities to start a new project phase. Main tasks for next days are:

• Upload information to google code project (http://code.google.com/p/informedproject/)
• Publish source code to SVN with the defined structure
• Linux re-installation in Condor
• System deployment to Condor
• To deal with server relocation (Telemedicin)

I created a new project in dotProject called Informed with these activities. Next meeting will be at September 17.

Relevance feedback

Well, this week I have been reading some papers about relevance feedback for image collection exploration and I find some interesting works in this address. The papers can be found in CiteULike (http://www.citeulike.org/user/camargoj). For me, the most interesting work was published by NGuyen at ACM Transactions on Multimedia last year (2008): "Optimization of Interactive Visual-Similarity-Based Search". Basically, authors propose a system that uses active learning for interacting with visualization of image collections. The build a summarization that is visualized to user, the user selects which images are relevant, the system applies a learning technique selection (SVMs), the system then retrieve new images according to the user feedback, the user repeat the process until target images are found. They did experiments with some image data sets and evaluate recall vs number of iterations.

I am working in Exploratory image collection search using a kernel-based relevance feedback model. I expect to have a more clear idea for sharing it with my advisor before the end of this week.

Manipulating a composed kernel space

These days, I am working on defining a model that allows to manipulate a kernel space, which is composed by individual kernels (one for image visual feature), such a way that the visualization can be modified by adjusting the weight of the kernels.

My research

Hi, in this blog I will be publishing advances of my research. The complete information about my research (publications, software, links, conferences, etc.) is at http://camargoj.googlepages.com.