rough estimation of angle and accuracy with certain angle error about rotnet HOT 3 CLOSED

Sorry for the late reply.

1. You can just restrict the number of outputs in the last layer. For example, you can have 36 outputs instead of 360 and then only rotate the images in multiples of 10. Then you simply map each output to a rotation angle (class 0 -> angle 0, class 1 -> angle 10, etc).
2. You can write a metric that outputs 1 if the angle is less than X degree and 0 otherwise.

from rotnet.

Thank you for your reply. Let me ask some other questions

then only rotate the images in multiples of 10

The problem is that we may have any rotation (not only in multiples of 10)
but we may have results that aren't very accurate (for example given that we need to track multiples of 10, when picture is rotated by 66 it should give 70 as a result. given that we have a metric where less then X degree angle error
Do we still rotate in multiples of 10 for training purposes or we don't need that?

Then you simply map each output to a rotation angle (class 0 -> angle 0, class 1 -> angle 10, etc).

How to do it to include it in y_pred?

from rotnet.

You can rotate the images with any degree and then generate the ground truth label by rounding the true angle to the closest multiple of 10. So if a picture is rotated 66, provide 70 as the ground truth label during training.

I'm not sure what you mean by your second question. You can just multiply the correct class by 10?

from rotnet.