Oscar Serra - Personal Home Page

I am a Ph. D. Student in the field of Artificial Vision in Robotics
(Control, vision and Robotics - Computer Science - Artificial Intelligence).

I belive that in order for a Cognitive Agent to have a robust perception (and consequently a robust interaction with the environment ), a fusion between several perception domains (vision, sound, temperature, compass, accelerometer...), as well as modes (cues), has to be done. I belive as well, as I observed in my first research year (2002-2003), that it is important to provide the agent with a certain amount of adaptation capabilities; we cannot pretend to think of every situation the agent is goint to face in the future: learning is key.

Neural Networks with Specific Architectures provide a natural solution to different adaptability problems (such as information compression, cause-effect learning, pattern classification, and a lot more) and are quite suitable for perception fusion. If we knew how to give to a robot enough learning capabilities, the cost of software development could be greately reduced. Bottom-up Emergence may be the key for future Intelligent Robots.

A particular model has attracted my attention, a neurobiologically inspired computer vision system, which at the time was among the best performant. After following this and other related works for more than two years now, I have specialized in feature extraction techniques. I am mostly interested in ways of rapidly generating overcomplete codes, which is believed to give extra selectivity to the original features, which in turn should allow for better object recognition performance, specially in clutter environments. Other authors have also found a way to relax some heuristics by learning the complex cell responses. My ultimate goal, for now, would be to construct a 3-layered architecture able to recognize objects in real time. Further improvements should take into account color and movement, and should mix several cues like optical flow, stereo vision segmentation, and probably more.

• Final Project (250K)

• Half term presentation (600K)

• Seminar report: Attention et Situation en Vision Artificielle (650K)

• Final report: Modélisation de Certains Aspets de la Mémoire Humaine à Travers una Approche Connexionniste Dynamique (15M)

• Final DEA presentation (2M)

• Neural Networks simulation software developed in C++: johnny-0.1-statique and johnny-0.2-dynamique

• Short explanation of the Itti and Koch Saliency Model (200K)

• Explaining "Cooperative Coevolution" of Genetic Algorithms and "Multiobjective Optimization" in Ensembling Neural Networks (350K)

• Improvement on the software in C++ and Qt: johnny-0.3-interface, johnny-0.4-qt and johnny-0.5-ga

• Invariant object recognition for a task-specific mobile robot

• Serre & Poggio Hierarchical Vision Model

• Hierarchical Model for Object Recognition

• Technical report on "Parallelization of Bio-Inspired Convolutional Networks for Object Recognition Using the GPU"

• Ph. D. Thesis Project

• GPU programming and CUDA comes into the play

• Overview on Nvidia CUDA

• A transparent library for Computer Vision acceleration in GPU (to be submitted soon)