Train with PRE_ORI : 'False'

[mahiratmis]

when I train the network with PRE_ORI : 'True' everything looks fine. I tried training the model with PRE_ORI : 'False'. However, during trainig where I save the initial predictions during epochs I observe that initial predictions have some kind of white layer on them. Do you have an idea of why it could be so ? Should it be like that or is something wrong? How can I address the problem ? Any help and feedback is really appreciated.

You can see the attached image for better illustration of the problem. The results are from a batch of images. Left picture is ground truth, middle picture is initial prediction and right picture is (ground truth - inital prediction).

[Royalvice]

You're right. When PRE_ORI: 'False', the visualization code during training is incorrect.

[mahiratmis]

Would it affect the inference ? If inference uses the results of the initial prediction the inference will also yield in bad results.

[mahiratmis]

The inference also produces results with white layer. Can you help adress the problem ?