Graphical processors programming for image processing and vision.

Recent advances in General Purpose computation on Graphics Processing Unit (GPU) leads to a computing power over the Teraflops available on a standard workstation, which is 15 times better than conventional CPU capabilities. This performance derives from a massively parallel architecture: standard GPU includes more than 400 computing cores (while classical multi-core CPU are limited to 8). This evolution implies a renewing of programming schemes, so as to fully benefit of the available ressources.

     
Figure 1 : Optical flow result on a full HD aerial video (1920 x 1080) taken over a canyon. The color map represents both flow norm and orientation, according to the pattern shown in the upper left corner. White regions stand for zero norm or homogeneous intensity region, where optical flow is not observable. Computing time is 25ms/frame.

In 2008 developments on GPU for image processing and computer vision have been launched in the Modeling and Information Processing department. This website is dedicated to some results of this activity, conducted in the EVS (Estimation and ViSion) research unit. Our first achievement is a highly efficient GPU code for optical flow estimation, based on an algorithm published in 2005 [1].