1, How did you get your data from Caltech Pedestrian Detection Benchmark? What is their corresponding relations?

Ans: We have trained one linear-svm to scan all caltech dataset images and extracted a lot of bounding boxes together their corresponding svm scores (See Section.4 in our paper). Those boxes will form the mat files we provided.

The code to transform image to mat files is very simple. There are three kinds of mat files. 

Suppose we have n images(in the Caltech Train data, n = 4250)

a, CaltechTestAllimBoxesBeforeNmsRsz3.mat

1*n cells with each cell for one image,

AllimBoxesBeforeNmsRsz{i}{j}.im is the image patch of the jth bounding box in the ith image.
AllimBoxesBeforeNmsRsz{i}{j}.score is cascade score of the jth bounding box in the ith image.

b, CaltechTestAllimBoxesBeforeNmsRszLabel3.mat

n*1 cells with each cell for one image

Labels{i}(j) is the label of the jth bounding box in the ith image.

If the jth bounding box is positive, Labels{i}(j)=1;
if the jth bounding box is negative, Labels{i}(j)=-1;

c, CaltechTestAllimBoxesBeforeNmsRszBox3.mat

1*n cells with each cell for one image.

Allpartboxes{i} is m*5 matrix with each row [x1 y1 x2 y2 score].

2, How to define the part filters?

We use 20 part filters and their corresponding positions are defined in cnnsetup3.m. 

We use four variables in cnnsetup3.m to define those positions, StartRow, EndRow, StartCol & EndCol. Those variables are 1*20 vectors, meaning 20 parts. Each column is one part filter.

You can find those part filters in Figure 3(a) in our paper (http://www.ee.cuhk.edu.hk/~xgwang/papers/ouyangWiccv13.pdf).

We suppose the whole pedestrian to be 15*5 size. Thus the ranage of StartRow & EndRow is [1,15], the range of StartCol & EndCol is [1,5]. 

For example, the first part filter (Row 1-3, Col 1-3) is the left-top head-shoulder part; the second part filter (Row 1-3, Col 3-5) is the right-top head-should part.

If you want to define your own part filters, you need to modify those four variables.