We have precomputed bounding boxes the detector run using a stride = 8x8, 
scaleratio = sqrt(sqrt(sqrt(2))), and the images are padded with 8px. 

For reference we have also computed the bouding boxes using the INRIA
Pedestrain Detector (as implemented in Dalal&Triggs,CVPR'05) run at the 
same scaleratio and a finer one 1.05. 

Run main.m to run the benchmarks using precomputed detections. It outputs 
the figure benchmark.png/fig. Inside benchmark_det.m turn on the DISPLAY
flag to see the detec
