PovertyMap Images

[niki-amini-naieni]

Hello,

I know the PovertyMap dataset includes images with eight different channels. As a result, I am wondering which channels the authors plot in the following image:

(p. 15 of original paper (link here: https://arxiv.org/pdf/2012.07421.pdf))

Thanks!

[sangmichaelxie]

The plotted images are the RGB channels without mean and std normalization, and then divided by the maximum pixel value of the entire image. Hope that helps!

…

On Tue, Mar 15, 2022 at 2:09 PM Niki Amini-Naieni ***@***.***> wrote: Hello, I know the PovertyMap dataset includes images with eight different channels. As a result, I am wondering which channels the authors plot in the following image: [image: image] <https://user-images.githubusercontent.com/61301874/158443524-747585cf-4b3f-4a41-860e-2ff162d63810.png> (p. 15 of original paper (link here: https://arxiv.org/pdf/2012.07421.pdf )) Thanks! — Reply to this email directly, view it on GitHub <#111>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB4J2SUQMJ3DWW3LIMXMJ53VADG6TANCNFSM5QZPYVNQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[niki-amini-naieni]

Are the RGB channels the first three (out of the eight) channels?

[sangmichaelxie]

Yes, the channel order is the order in which it appears in the MEAN and STD: BLUE, GREEN, RED, SWIR1, SWIR2, TEMP1, NIR, DMSP/VIIRS

…

On Wed, Mar 16, 2022 at 4:49 AM Niki Amini-Naieni ***@***.***> wrote: Are the RGB channels the first three (out of the eight) channels? — Reply to this email directly, view it on GitHub <#111 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB4J2SVLXI4IUUYUO2I4UULVAGOBRANCNFSM5QZPYVNQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you commented.Message ID: ***@***.***>

[niki-amini-naieni]

Thank you. I see 9 means and standard deviations in the code shown below:

How do you combine the last two channels to produce a total of eight channels and how do you incorporate the means and standard deviations of the last two channels into the construction of the eighth channel? Also, how do you justify using the means and standard deviations of all the images (including images in the test set) to normalize images in the training and validation sets?

[niki-amini-naieni]

I see that the nightlight data comes from two different sources and images from each source were normalized separately (i.e., each with a separate mean and standard deviation). I am including the following image and link for reference in case anyone else has the same question in the future.

https://www.nature.com/articles/s41467-020-16185-w.pdf

[sangmichaelxie]

That's right, before 2012 the nightlights are from DMSP and for 2012 and after it is from VIIRS. We just followed the normalization process from Yeh et al. (https://www.nature.com/articles/s41467-020-16185-w). The means and std shouldn't be test set statistics.

…

On Thu, Mar 17, 2022 at 7:47 AM Niki Amini-Naieni ***@***.***> wrote: I see that the nightlight data comes from two different sources and images from each source were normalized separately (i.e., each with a separate mean and standard deviation). — Reply to this email directly, view it on GitHub <#111 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB4J2SXJ4A4NOEWHCWEMDVTVAMLWJANCNFSM5QZPYVNQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you commented.Message ID: ***@***.***>

[niki-amini-naieni]

Thank you for the clarification! Perhaps I am misinterpreting the comment below. Could you please let me know what this comment implies?

(https://github.com/p-lambda/wilds/blob/main/dataset_preprocessing/poverty/dataset_constants.py)

[sangmichaelxie]

Actually, you're right - again, we are following Yeh et al., but I believe the reason this was done is because we test on 5 different folds of the data, and 1) the average evaluation metric over folds is already a function of all the folds, 2) convenience, but I agree that perhaps a better way to do it is to compute a fold-specific mean/std using each training set.

[niki-amini-naieni]

Thank you for the confirmation! I completely understand (: