Pytorch_Retinaface
Pytorch_Retinaface copied to clipboard
biubug6
人脸姿态估计+人脸检测+五点检测 1个网络实现
backbone=resnet50
@qinxianyuzi 1.采用hopenet模型对widerface数据集进行姿态预标注 2.参考retinaface口罩分支 重新训练，即可
采用hopenet模型对widerface数据集进行姿态预标注
你的口罩的分支怎么训练的，可以分享一下idea么？是增加一个branch专门做口罩的property输出，还是num_class改成人脸，口罩两个类型？
@qinxianyuzi 我采用的是hopenet的loss函数，没有修改； @stevenluzheng123456 基于widerface的数据进行操作的
moscaic的操作会造成人脸的distortion，这个会对性能有影响么？
On Tue, Feb 2, 2021 at 2:11 AM gengyanlei [email protected] wrote:
加1个类似口罩的分支，用于姿态估计，并且加1个mosaic操作！
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/biubug6/Pytorch_Retinaface/issues/145#issuecomment-771295501, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEHBG4GDGE47UCBM5NFBDATS45NNDANCNFSM4UWEK5YQ .
--
[image: Employee profile] Steven Lu Senior Researcher Research Team
M +447712148438 E [email protected] [email protected] W www.anyvision.co
Tel: 02890737904
@stevenluzheng123456 我舍弃了一部分操作，只在batch=8时精度提升明显，其它均下降，原因是因为mosaic操作，没有精确设计！
@stevenluzheng123456 @qinxianyuzi 测试代码，权重都在这里了 https://github.com/gengyanlei/Pytorch_Retinaface
Yes, mosaic was performing well , especially for small face detection, but my mosaic method is slightly different, I didn't jeopardize the ratio of image, I concatenate 20-100 small single images into a big image.
On Sat, Jun 5, 2021 at 1:18 PM alicera @.***> wrote:
How about your experiment for mosaic?
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/biubug6/Pytorch_Retinaface/issues/145#issuecomment-855232001, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEHBG4DBHROWMQW7Q4LDSP3TRIIYXANCNFSM4UWEK5YQ .
--
[image: Employee profile] Steven Lu Senior Researcher Research Team
M +447712148438 E @.*** @.***> W www.anyvision.co
Tel: 02890737904
@alicera
- 提点1.2个百分点，由于我是人脸合并后降低，然后采用mosaic提升了，从93.4%->94.43%，原始为94.8%
- mosaic我没有设置合理，应该还可以提升，这个需要一些经验。