Normalization

[ryanxingql]

Hi Lee, I notice here we store the estimated map after converting it into an image. This may introduce normalization. I am wondering what is the network output range? What if the output ranges outside [0, 1]? Thank you!

[codeslake]

Hi, @ryanxingql.

The predicted defocus map is in the range [0, 1], so the scipy.misc.toimage() function does not normalize the defocus map as we provided the appropriate range. Note that the ground truth defocus maps are also in the same range between 0 and 1.

We've provided sigma_map to get the actual sigma value.

[ryanxingql]

Thank you! And also, what is the use of the sigma map? Is it for deblurring?

[codeslake]

Thank you! And also, what is the use of the sigma map? Is it for deblurring?

Yes!

[ryanxingql]

Thank you so much ^ ^

[ryanxingql]

Hi @codeslake, I would like to ask you a question.

My understanding: Our gray-scale defocus maps are ranging from 0 to 255. Therefore, pixels with 0-values are focused, while the maximal available defocus value is 255 since the maximal COC for training is 15. If a object in test images reaches out of this COC, its defocus value is still 255. In this situation, the output is somehow incorrect.

Am I right? Thank you.

[codeslake]

Yes, you are right. DMENet will not be able to accurately predict the COC values of a test image, if the COC values are larger than the maximum COC used to create the training set.

[ryanxingql]

@codeslake Thank you!