Number of training images for HOG people detection
Hi all. Can any one give his/her experience on training SVM for people detection using HOG? I don't mean how to do so, but I mean how many images for training is sufficient for successful detection: How many positive and so negative images.
If we consider INRIA person set that was used in Dalal-Triggs' paper, I went through the folders and files and I got confused with the following:
The folders are organised in a way that seems incorrect for me, or perhaps I understand the term 'train' and 'test' wrongly. There is a folder named '70X134H96' and contains a subfolder 'test' then 'pos'. Is this folder meant to be positive images for training? If no, then which folder?
How is it to train negative images? Do I crop randomly from the 'neg' folder and load it into the SVM as a negative image?
Thank you :)