Open hungsing92 opened 1 year ago
Thanks for your appreciation. In some samples, the registration of side face may fail since the target mesh does not have that region. Therefore, we weight the l2 loss of each vertex, that this, high weights for facial features, medium weights for cheek, and low weights for other regions (including ears, neck, where registration easily fail).
Can you provide train code? Thanks very much!!!
首先非常感谢您的工作!
有少许问题想向您请教一下。我可视化了一下FG3D数据集的数据,发现质量不是特别好,脸部很多疙瘩,有些侧脸不完整和变形。请问这个是正常的吗?对于侧脸不完整这块论文里貌似没看到loss的处理方式。另外请问数据集使用的人脸模型是bfm吗?是哪个版本的bfm?另外可以提供训练代码吗?
非常感谢!