All of the code below was developed for my specific desktop system (Ubuntu 11.04, Python 2.7) and the software packages mentioned on the references page. You will need to preprocess the LFPW data (converting all images to JPG format and removing corrupted downloads from broken links). Then you will need to modify the code to reflect the file system structure you use for saving the data. All code is released under the GNU Public License and is offered as-is with no warranty of any kind.
The file linked below contains the GPU and CPU source code for the HoG implementations, source code for the Dalal and Triggs HoG extractor, scripts for building single-fiducial SVMs and poselet SVMs, a script for downloading the LFPW data set once you have the publicly available LFPW CSV files, a script for timing the CPU vs. GPU versions, and test images for the timing analysis.