Per-pixel confidence loss question

[phongnhhn92]

Hello,
In the equation (5) in your paper, there is a element-wise square operator like this:

I believe this loss function was defined in this piece of code
l1_loss += tf.reduce_mean(loss_map * normalized_mask) * current_weight / \ (int(img.get_shape()[1]) * int(img.get_shape()[2])) * regularizer_weight

I wonder it should be like this
l1_loss += tf.reduce_mean((loss_map**2) * normalized_mask) * current_weight / \ (int(img.get_shape()[1]) * int(img.get_shape()[2])) * regularizer_weight

Can u verify it is correct ? If your code is correct then I am sorry for my question. I just confused when I see the loss like that. Thanks !

[shaohua0116]

The code uses the l1 norm (as the loss map takes absolute values), which performs similarly to the element-wise square operator in the paper. To use the one mentioned in the paper, you can change the code to tf.reduce_mean(loss_map * normalized_mask) ** 2.