PFLD
PFLD copied to clipboard
guoqiangqi
3D points
hello，where can find the standard 3D points define， any references can provide，thanks
Take this web as a reference "https://github.com/lincolnhard/head-pose-estimation"
@guoqiangqi 请教一下，backbone网络的输出是196个[0, 1]区间的小数，换算为实际坐标的方法是：分别对应乘以[height, weidth]；我现在输出取辅助网络auxiliary的输出来算姿态，得到的输出是pitch，yaw， roll 三个小数，范围在[-1, 1]之间，那么我要怎么把三个小数转化为具体的角度呢，请指点一下。
分支网络直接预测角度，角度的 GT 根据 3D标准模型的关键点坐标和人脸的标注坐标就可以计算得到。
通过训练数据的关键点和3D标准模型来计算姿态这部分是看了。不好意思，我没表达清楚，我现在是用训练好的模型来预测姿态，好像有2个实现的方法： 1.跑inference直接取分支网络的输出，但是这个输出是3个小数，范围在[-1, 1]之间，不知道怎么转换为角度GT. 2.通过跑inference 主网络得到关键点，再和3D标准模型来计算角度GT.
我现在卡在问题1了，没弄清楚。当然方法2是可以解决的，就是时间成本问题。
(1)预测的角度如何转化为角度GT，取决于你在计算euler angles loss时预测值与GT值的对应关系，而且这个如果目标是姿态角度，建议针对角度预测网络训练，而不是关键点检测。 （2）如果目标是检测姿态，用主干网络检测关键点再计算角度，感觉会比较慢的
Take this web as a reference "https://github.com/lincolnhard/head-pose-estimation"
ok，thanks for your reply， but i'm confused about the input size of images, because i found the calculated pose is different if the input size different,(i padding the image by opencv copyMakeBorder function）,so, how to handle it?
Take this web as a reference "https://github.com/lincolnhard/head-pose-estimation"
ok，thanks for your reply， but i'm confused about the input size of images, because i found the calculated pose is different if the input size different,(i padding the image by opencv copyMakeBorder function）,so, how to handle it?
What do you mean different input size,the different input size corresponding to different eulers angles?The input of the PFLD or input of the euler angles demo,you can try this demo form https://github.com/guoqiangqi/euler_angles
Take this web as a reference "https://github.com/lincolnhard/head-pose-estimation" ok，thanks for your reply， but i'm confused about the input size of images, because i found the calculated pose is different if the input size different,(i padding the image by opencv copyMakeBorder function）,so, how to handle it?
What do you mean different input size,the different input size corresponding to different eulers angles?The input of the PFLD or input of the euler angles demo,you can try this demo form https://github.com/guoqiangqi/euler_angles
sorry， I may not express clearly， just about calculate euler angles by using 2D landmarks and 3D points, not PFLD net input, because the input of image in http://aifi.isr.uc.pt/Downloads/OpenGL/glAnthropometric3DModel.cpp is 800x600, so, whether the 3D points correspond a standard input image size?
Take this web as a reference "https://github.com/lincolnhard/head-pose-estimation" ok，thanks for your reply， but i'm confused about the input size of images, because i found the calculated pose is different if the input size different,(i padding the image by opencv copyMakeBorder function）,so, how to handle it?
What do you mean different input size,the different input size corresponding to different eulers angles?The input of the PFLD or input of the euler angles demo,you can try this demo form https://github.com/guoqiangqi/euler_angles
sorry， I may not express clearly， just about calculate euler angles by using 2D landmarks and 3D points, not PFLD net input, because the input of image in http://aifi.isr.uc.pt/Downloads/OpenGL/glAnthropometric3DModel.cpp is 800x600, so, whether the 3D points correspond a standard input image size?
Reference to https://blog.csdn.net/cdknight_happy/article/details/79975060