Parallelization of image processing and computer vision algorithms
Parallel processing in multi-processor systems
Design and development of algorithms tailored to parallel architectures (multi-core and GPGPU):
- Uniprocessor performance is growing ever more slowly in the last decade.
- Video sources are growing in quality (and therefore size) faster than uniprocessor performance.-
- Partial solution: multicore CPUs, but..
- Traditional sequential CPUs are not optimized for numeric computing throughput.
- Consumer-grade stream processors, aka GPUs, are designed for high throughput and energy efficiency in numeric computing.
- Data-parallel portions of computer vision algorithms can run 10x faster in these devices and help achieve real-time operation.
- Fast-evolving technology, still in its infancy.
- The programmer must take into account many architecture-specific details to get good performance.
