How can I properly test my accuracy using the INRIA data set

Question

I am working on a pedestrian detection algorithm and I want to test its accuracy using the INRIA data set. The dataset has two sets of test folders that contain positive images. One has cropped versions with pedestrians located in the middle within a 64 x 128 pixel window. The other folder contains the uncropped version with annotation files for the bounding boxes of each pedestrian. If I use the uncropped version with annotation files, how do I say that my detection box is correct given that it may not have perfectly the same value as the bounding box value in the annotation file but still accurately detect the pedestrian. If I used the cropped version, will I still scan the image or simply extract the middle 64 x 128 window containing the pedestrian. I aim to accurately compare my results with the work of Dalal and Triggs.

Answer 1

A successful detection can be defined by atleast a 50% overlap of the detected bounding box by your method and that provided in the annotation file.

reference: Google Scholar http://scholar.google.com/scholar?q=%22at+least+50%25+overlap%22+object+detection