 |
PhD candidate (2007 - exp. 2010) Tel.: +33 (0)6 08 84 79 84 email : trongtonfr at yahoo dot fr blog: http://trongton.free.fr/blog |
I am interested in pattern analisys techniques and multimedia retrieval (such as image and text). My objective is to build a solid experience on this emerging subject in order to apply to an online search engine.
Peoples:
I am working closely with Dr. Philippe Mulhem - CNRS researcher at MRIM - LIG and Dr. Lim Joo Hwee - Singapore Director of IPAL lab on my thesis. I have also oppotunities working with Jean-Pierre Chevalet, Nicolas Maillot and other collegues at Institut for inforcomm Research (I2R) - Singapore.
Research Topic:
Image tagging, the problem of automatic annotating of an unseen image with some keywords to help an image retrieval system, has been intensively studied in the last decade. In the literature, most of current techniques are based on the “bag-of-word” model. This method is simple as each visual entity (i.e. region, patch, interest point) is processed individually, therefore important information on the image structure is lost.
Relational graphs are useful models to describe image structure. For example, images related to the concept “cityscape” will usually have the visual entities “building”, “sky”, “road” and “car” (these are also the keywords that annotate the images). A rough structure persists throughout these conceptually similar images, i.e. the sky is on top in the background, the building is the center of the image, while roads and cars are at the bottom. These separate entities in the images do not appear in arbitrary positions in the images. One can imagine that if structural relations between individual entities are considered, a model for a particular concept can be built more effectively.
In this work, we propose to apply relational graphs in a manner that is more suited for concept modeling. Specifically, we estimate one relational graph for one concept. To do this, we merge the individual relational graphs of each image belonging to a common concept to obtain a signature relational graph for that concept. This should allow us to focus on the common attributes contained in conceptually similar images, and also by diminishing the effects of spurious visual entities that do not contribute towards describing the concept, thus obtaining a more accurate conceptual model.
Projects:
GRASP-IT project (2008-2010), Franco-Singaporean scientific colaboration
Publications:
- Trong-Ton Pham, Jean-Pierre Chevalet and Joo-Hwee Lim. "Une fusion multi-modalités par un modèle de sémantique latente". Conférence en Recherche d'Information et Applications (CORIA 2008), Trégastel, France, Mars, 2008 [to appear]
- Trong-Ton Pham, Nicolas Maillot, Joo Hwee Lim and Jean Pierre Chevallet. "Latent Semantic Fusion Model for Image Retrieval and Annotation". ACM Sixteenth Conference on Information and Knowledge Management (CIKM 2007), Lisboa, Portugal , November, 2007. [PDF]
Links:
My CV in English [pdf] and in French [pdf]
Some of my photos can (lah!) be found online here and there
Hobbies:
I am free for any discussion, football match, online game, photography or even a beer ;)