cannot figure out how to multiply or divide a channel by a scalar in order to renormalize it

[Chouffe]

Using origami with MXNet, I need to be able to renormalize a channel color by a scalar before feeding the image into my networks. I looked up examples in Python and they often rely on numpy to perform such an operation: http://akash0x53.github.io/blog/2013/04/29/RGB-Normalization/

Is there a way to achieve this with origami? I have tried using cv/multiply and cv/divide without success...

[hellonico]

Yes, origami has a function for that but getting terrible results due to black pixels.

In the formula from the link you sent, I don't really get the divide by 0 in case the sum of R+G+B on a black pixel (0,0,0) gives a sum of 0.

Is it suppose to be 0 ? or something else ?

[Chouffe]

Yes the example above is broken for black pixels. And I also do not care about the scaling factor in his example. In my case, total would be the standard deviation of the color channel values across the image dataset. So it will never be black.

[hellonico]

When the channels are independent you can use the following:

(deftest multiply
    (let [m (cv/new-mat 2 2 cv/CV_8UC3 (cv/new-scalar 10 20 30))]
     (cv/multiply! m (u/matrix-to-mat-of-double [[1.0 0.0 0.0]]))
     (is (= (cv/->string m)
        [[10 0 0 10 0 0]
         [10 0 0 10 0 0]]))))

(deftest divide
    (let [m (cv/new-mat 2 2 cv/CV_8UC3 (cv/new-scalar 10 20 30))]
     (cv/multiply! m (u/matrix-to-mat-of-double [[1/5 1/10 1/30]]))
     (is (= (cv/->string m)
        [[2 2 1 2 2 1]
         [2 2 1 2 2 1]]))))

(deftest divide-bikkuri 
        (let [m (cv/new-mat 2 2 cv/CV_8UC3 (cv/new-scalar 10 20 30))]
     (cv/divide! m (cv/new-scalar 10.0 20.0 30.0))
     (is (= (cv/->string m)
        [[1 1 1 1 1 1]
         [1 1 1 1 1 1]]))))

[hellonico]

Added related example to channel transformation using functions or matrixes: https://github.com/hellonico/origami/blob/a2c855a1138dd61cebc8303b7008c5aec370dcec/test/origami/channels_test.clj