Performance optimization of color segmentation algorithms

Video object segmentation is a key technique for future multimedia systems with content-based access to visual information. Various techniques for performance optimization of' video object segmentation algorithms have been examined. Special requirements arise from the irregular pixel addressing scheme, common to most segmentation algorithms. Instruction profiling has shown that pixel addressing operations are clearly dominant over pixel processing. Therefore a structured pixel addressing scheme was selected for further performance optimization. This addressing scheme was used in a software implementation of the algorithm, as well as in a hardware architecture of a configurable coprocessor. In order to evaluate the potential of hardware acceleration for video object segmentation, both implementations have been compared with respect to throughput and a metric for implementation cost. The benefit of the hardware architecture over the software implementation was a factor of 25 for throughput and a factor of 150 for the metric. This is due to optimized pixel addressing in the color segmentation algorithm.