Re-train model without shadow mask

[hnrna]

You can use an existing shadow detector to pre-predicted a shadow mask, then input it into the ShadowFormer. Or you can re-train the model without the shadow mask.

Originally posted by @GuoLanqing in https://github.com/GuoLanqing/ShadowFormer/issues/7#issuecomment-1445648681

Hello, have you experimented with retraining the model without the shadow mask? If you have tried, is the model trained in the above way still able to achieve the effect of de-shadowing? And how do the evaluation metrics perform?

[wmj183363206]

No, I still use the dataset with shadow mark. There is a 'xm' in forward function, which I think is the required shadow mask.

[GuoLanqing]

You can use a pre-trained shadow detector if there is no GT shadow mask. Alternatively, you can train without the shadow mask, but you need to remove the SIA module.

[hnrna]

Thank you for your suggestion, I will try to modify the model.