Low-Level Image Features for Real-Time Object Detection
Adam Herout, Pavel Zemčík, Michal Hradiš, Roman Juránek, Jiří Havel, Radovan Jošth, Lukáš Polok · InTech eBooks · 2010
This contribution presents the Local Rank Differences/Patterns low-level image feature extractor and its efficient implementations on several hardware architectures. This image feature set was not only developed to provide equal classification performance as its stateof-the-art alternatives, but to be executed much more efficiently in hardware implementations either programmable hardware (FPGA) or custom specialized chips (ASIC). However, the feature set performs well also on more conventional platforms based on processors. The measurements given in section 4 show that the speed achieved by using Local Rank Functions namely Local Rank Differences (the special case) is interesting. The baseline implementation outperformed the state-of-the-art Haar wavelets (especially in case of higher resolutions), and the hardware-accelerated implementations speeded-up the baseline LRD implementations more than by order of magnitude. Measurements show that the performance on the GPU's is equal for CUDA and GLSL programming. Considering that CUDA is much more intuitive and compatible to standard C language programming, the conclusion can be drawn that CUDA (or possibly OpenCL in near future) is a good selection for exploiting graphics hardware for non-rendering tasks, such as object detection. Future work should include exploiting more general Local Rank Functions and conducting further investigation of the combinations of the features with more traditional ones.