HOG cell visualization

Moderator: Computer Vision

SebFreutel
Computerversteher
Computerversteher
Beiträge: 317
Registriert: 30. Okt 2006 21:54

HOG cell visualization

Beitrag von SebFreutel »

I am a bit confused about the images on slide 17 of the hog/svm lecture slides:
As I understand, the third image shows the cells for an average positive example. The most visible (which I assume to be the heighest weighted) lines in the cells are the ones corresponding to the contours of a person.
But shouldn't they be perpendicular to the contours as the cells contain a histogram of gradients?

diane
Windoof-User
Windoof-User
Beiträge: 25
Registriert: 21. Jun 2009 17:33

Re: HOG cell visualization

Beitrag von diane »

This is a good remark. Contours and gradient vectors are suppose to be orthogonal. And that's the visualization we are using for the exercises. But in the case of slide 17, a different visualization scheme has been chosen: the gradients are presented rotated by 90°, so the contours appear in a more visible manner. That is a different convention which is sometimes used in papers, and PhD thesis, likes for instance Dalal's thesis who proposed the HOG descriptor.

SebFreutel
Computerversteher
Computerversteher
Beiträge: 317
Registriert: 30. Okt 2006 21:54

Re: HOG cell visualization

Beitrag von SebFreutel »

Thanks for the clarification

Antworten

Zurück zu „Computer Vision“