why predict foreground and alpha together, not alpha only?

[Stick-To]

I think src * alpha = foreground

[PeterL1n]

No, it's not.

src = pha * fgr + (1 - pha) * bgr

The foreground will be exactly the same as src only for pixels where alpha is 1. But for semi-transparent regions, you need to predict the actual foreground.

[Stick-To]

got it, thanks

[zhanghongyong123456]

No, it's not.

src = pha * fgr + (1 - pha) * bgr

The foreground will be exactly the same as src only for pixels where alpha is 1. But for semi-transparent regions, you need to predict the actual foreground.

您好，我数据集只有合成图以及对应的alpha，我该如何进行修改呢，1.必须要对合成图像的 前景、背景分离，这一步该怎么操作，是否可以根据src = pha fgr + (1 - pha) bgr，将bgr设置为白色或者黑色进行合成得到像你给的数据集样子，这样操作正确吗；2.或者将训练代码的预测前景代码进行注释，只进行alpha预测，不预测前景对模型最终的性能影响大不大。