1 | initial version |
if you cannot guarantee, that folks are looking stright into the camera, try the following:
use 2 CascadeClassifiers, one with a frontal face detector, and the other with the profile detector
then, you have to do 3 detections, 1st with the frontal, 2nd with the profile, and a 3rd profile detection with the flipped image.
"While i will try to train my own cascade"
i have some doubts, that you will be able to get better results. (you'd need a ton of data for this)
2 | No.2 Revision |
if you cannot guarantee, that folks are looking stright into the camera, try the following:
use 2 CascadeClassifiers, one with a frontal face detector, and the other with the profile detector
then, you have to do 3 detections, 1st with the frontal, 2nd with the profile, and a 3rd profile detection with the flipped image.
then use IOU to filter out duplicate detections
"While i will try to train my own cascade"
i have some doubts, that you will be able to get better results. (you'd need a ton of data for this)