3D rendering onto stereoscopic VR goggles requires that the depth of the scene is known at a resolution corresponding to the one of the final 3D images to be rendered. Active depth sensors (e.g. Kinect) typically do not provide such high-resolution depth maps, therefore stereo/multi-camera matching techniques are often preferred.
This project/Master thesis aims at porting a C/C++ high-quality depth estimation algorithm from the worldwide MPEG-I standardization committee onto an OpenCL, embedded device (e.g. Jetson TK1, DragonBoard, FPGA, hybrid platform, etc).
The advantage to have OpenCL code is that parallelization, even towards embedded graphics processors and FPGA devices (automatic translation from OpenCL to VHDL) is easily supported. Nevertheless, the specificities of the embedded platform should be considered for achieving best performances.

Prerequisites:

C/C++, mastering OpenCL and/or having followed INFO-H-503 (CUDA) is a plus.

Keywords:

Computer vision, Software development.

Promotor:

Prof. Gauthier Lafruit

Contact:

Gauthier Lafruit gauthier.lafruit@ulb.ac.be,

Daniele Bonatto daniele.bonatto@ulb.ac.be