Gradiant researchers attend summer schools in Pattern Recognition and Computer Vision

Antonio Torralba professor often collaborates with MIT

Gradiant | Redaction

The research lines in which Gradiant develops its activities are in constant evolution. It is imperative, then, for Gradiant researchers to be aware of the advance of science and technology in the areas in which the Centre works, assessing both its possible use in current projects and the launching of new initiatives. Besides technology watch, attendance to courses, seminars and conferences is a practice that Gradiant has been promoting since the beginning of its activities, four years ago. In the following, we show two recent examples.

Last June, several researchers from Gradiant attended the AERFAI Summer School 2012. For those who are interested, the videos of some of the lectures will be soon available in the web page of Campus do Mar, thanks to UvigoTV, the GTM group of the University of Vigo and Campus do Mar.

As well, from 2 to 6 of July Daniel Pereira attended a seminar on “Visual Scenes and Object Recognition in Computer Vision” that was held at University Carlos III, Madrid. The seminar, which was part of the Máster Interuniversitario en Multimedia y Comunicaciones schedule, was given by Antonio Torralba (in the photography) associate professor at MIT, who reviewed the key elements that compose the object detection techniques most used nowadays (support vector machines, boosting, HOG and Haar descriptors, etc.).

Moreover, professor Torralba introduced interesting concepts such as transfer learning among object classes which share common features when there are not enough training examples; or scene recognition through global features (GIST) and context analysis, which can be used to determine the probability of co-occurrence of certain object classes depending on the kind of scene.

He addressed also an important, and sometimes ignored, problem which is the presence of biases in training image datasets. This can make an object detector to have a poor generalization capacity which, in the end, leads to drastic performance degradation in realistic scenarios (for example, a car detector trained only with side views will not behave properly in a scenario where vehicles are seen from other angles).

Attendance to this kind of events contributes to improve our work at Gradiant, adding new knowledge to several research lines like Human Sensing and Advanced Video Analysis, among others.