class_balanced_cross_entropy_loss_theoretical

[bhack]

Do you have any more info why class_balanced_cross_entropy_loss_theoretical didn't work so well?

[scaelles]

It didn't work because there is some numeric instability in the training with the logarithmic function. You can find more information in the TensorFlow inplementation without the class balancing.

[bhack]

Ok this passages remove also sigmoid(). Why you have used:

labels = tf.cast(tf.greater(label, 0.5), tf.float32)

[scaelles]

I used this to make sure that the labels are binary with 0 or 1 values.

[bhack]

Yes I understand this but Davis png annotation are not already 0 or 255?

[scaelles]

Yes, DAVIS has these values, but not all databases do. So I wrote this to enforce that the value of the labels is always correct.

[bhack]

Thanks

[j50888]

Hi, I have a question. Why the line 800 in the osvos.py use 162.0/255.0 instead of 0.5? res_np = res.astype(np.float32)[0, :, :, 0] > 162.0/255.0

[scaelles]

We use this threshold instead of 0.5 because it's the optimum. However, the results using 0.5 are almost the same.

[j50888]

Thanks!