尊敬的作者,
我在复现您的《Structure Destruction and Content Combination for Face Anti-Spoofing 》文章中的DCN模型时,采用OULUNPU Protocol1数据集进行训练,并使用文章中提到的《Single Image Reflection Removal with Perceptual Losses》工作来生成Reflection Map。但复现过程中,发现train_Reflect损失函数下降特别快,100epoch后ACER维持在0.4左右,且THER为万分位。
我也尝试使用深度图作为label复现您的工作,调整最终probs = 1 - probs,可以收敛。因此求教是否是Reflection Map生成时需要做一些额外操作,已达到您的效果?
期待得到您的回复!
尊敬的作者,
我在复现您的《Structure Destruction and Content Combination for Face Anti-Spoofing 》文章中的DCN模型时,采用OULUNPU Protocol1数据集进行训练,并使用文章中提到的《Single Image Reflection Removal with Perceptual Losses》工作来生成Reflection Map。但复现过程中,发现train_Reflect损失函数下降特别快,100epoch后ACER维持在0.4左右,且THER为万分位。
我也尝试使用深度图作为label复现您的工作,调整最终probs = 1 - probs,可以收敛。因此求教是否是Reflection Map生成时需要做一些额外操作,已达到您的效果?
期待得到您的回复!