Training Problem
YongtaoGe opened this issue · comments
Hi, @Jack-CV. It is really a nice work to detect the large pose face. Can you explain how to get the rotate angular label of face to supervise the trainning. As I know, WIDER FACE don't have the facial landmark. Do you use one of the facial landmark detect algorithms to get the facial landmark and then calculate the facial angular or just label all the faces manually?
Besides, what do you mean that "we rotate the initial training images uniformly in the range of..." in section 2.3 and 2.4 of the paper? As a beginner, I am confused of the statement.
@geyongtao Thanks for your interest in our work. WIDER FACE don't have the facial landmark or rotate angular. We use the images only containing upright faces as our training images(WIDER FACE provide the face attribute labels, e.g. pose), and rotate them to the rotation-in-plane(RIP) angle we want, the rotation angle is the rotate angular label.
The range of input faces' RIP angle in the first stage is [-180, 180], so we need to rotate the training images in the range of [-180, 180]. For the second and third stage, [-90, 90] and [-45, 45] respectively.
got it