3 ms·

One other problem is that "real" black and white images don't come from the same distribution as greyscaled images. The training data is just regular photos tha

One other problem is that "real" black and white images don't come from the same distribution as greyscaled images. The training data is just regular photos that have had all the color channels combined and averaged together. Sometimes it's a weighted average, like green is given more weight than blue, because human eyes see green better. However real black and white photos represent the actual intensity of the light waves. Black and white photos might be sensitive to light humans can't even see, or weight different "colors" very differently. Even two colors that appear the same to the human eye might be different intensities and frequencies, and so produce a different black and white photo. I don't know how much of an effect this would have. But real black and white photos definitely look different than greyscaled color photos, and it might lead the NN to guessing the wrong color information.

derefr 11y ago
Interesting point; the "perceptual brightness" formula used in grayscaling calculations is usually this: sqrt(0.299r² + 0.587g² + 0.114b²) Real "photosensitizing" chemicals, on the other hand, have spectral response curves, like this one[1] for the silver iodide used in daguerrotypes. (The formula above just being an approximation of the summed spectral response curves of the three cones in the eye.) There's a big difference, like you said, but it's not an impossible problem; we should be able to create a greyscaling formula for any photographic process with a known response curve. (And honestly, I have no reason to think this hasn't already been done in the creation of an Instagram filter at some point.) [1] https://books.google.ca/books?id=9IpaIAcgthQC&lpg=PA119&ots=zEWTsY30T3&dq=silver%20iodide%20photosensitivity%20spectrum&pg=PA120#v=onepage&q=silver%20iodide%20photosensitivity%20spectrum&f=false https://books.google.ca/books?id=9IpaIAcgthQC&lpg=PA119&ots=...
- Houshalter 11y ago
  Well not necessarily. If the response curves are different than that of a digital camera, then some information will be lost. In fact the response curves are different even for different cameras. But you could probably get a closer approximation using a better formula.
jrockway 11y ago
I've done a little bit of experimentation with this, comparing real B&W film to various software's conversions to B&W. Most interesting was a case where I took a picture of the NYC subway diagram with a red filter. I could never make the RGB data from the digital camera look like my original negative, even when shooting through the same red filter. (Which should not be necessary, if you want to block blue and green, you should be able to do that in software. Lightroom, at least, did not let me make a matching image.) More subjectively, I think most digital pictures converted to B&W look kind of dull, whereas actual film looks very exciting to me. I haven't done any detailed research into this, but I'm not 100% convinced that collecting luminance through red, green, and blue filters can capture all the data that panchromatic B&W film captures. (Even more of a tangent, one of the joys of B&W photography is that you can outright lie about colors and the photo still works. Try a red filter and watch the blue sky become black!)
jacobolus 11y ago
> However real black and white photos represent the actual intensity of the light waves No. Black and white negatives represent the intensity of the light (possibly with some lens filters in between) multiplied by some light frequency sensitivity distribution of the film, under some nonlinear time/intensity response function composed with another nonlinear time/chemical strength/exposure function from the chemical development of the negative. Then to get from negative to print, black and white photos undergo another pair of multiparameter nonlinear functions (representing the exposure and chemical development process), possibly including intentional manipulation by a human operator. Depending on the film, filters, development process, and printing techniques used, this can be relatively close or quite far from a naively converted 3-channel color picture. For details cf. for example Ansel Adams’s books The Negative and The Print.

derefr 11y ago

Interesting point; the "perceptual brightness" formula used in grayscaling calculations is usually this: sqrt(0.299r² + 0.587g² + 0.114b²) Real "photosensitizing" chemicals, on the other hand, have spectral response curves, like this one[1] for the silver iodide used in daguerrotypes. (The formula above just being an approximation of the summed spectral response curves of the three cones in the eye.) There's a big difference, like you said, but it's not an impossible problem; we should be able to create a greyscaling formula for any photographic process with a known response curve. (And honestly, I have no reason to think this hasn't already been done in the creation of an Instagram filter at some point.) [1] https://books.google.ca/books?id=9IpaIAcgthQC&lpg=PA119&ots=zEWTsY30T3&dq=silver%20iodide%20photosensitivity%20spectrum&pg=PA120#v=onepage&q=silver%20iodide%20photosensitivity%20spectrum&f=false https://books.google.ca/books?id=9IpaIAcgthQC&lpg=PA119&ots=...

Houshalter 11y ago
Well not necessarily. If the response curves are different than that of a digital camera, then some information will be lost. In fact the response curves are different even for different cameras. But you could probably get a closer approximation using a better formula.

jrockway 11y ago

I've done a little bit of experimentation with this, comparing real B&W film to various software's conversions to B&W. Most interesting was a case where I took a picture of the NYC subway diagram with a red filter. I could never make the RGB data from the digital camera look like my original negative, even when shooting through the same red filter. (Which should not be necessary, if you want to block blue and green, you should be able to do that in software. Lightroom, at least, did not let me make a matching image.) More subjectively, I think most digital pictures converted to B&W look kind of dull, whereas actual film looks very exciting to me. I haven't done any detailed research into this, but I'm not 100% convinced that collecting luminance through red, green, and blue filters can capture all the data that panchromatic B&W film captures. (Even more of a tangent, one of the joys of B&W photography is that you can outright lie about colors and the photo still works. Try a red filter and watch the blue sky become black!)

jacobolus 11y ago

> However real black and white photos represent the actual intensity of the light waves No. Black and white negatives represent the intensity of the light (possibly with some lens filters in between) multiplied by some light frequency sensitivity distribution of the film, under some nonlinear time/intensity response function composed with another nonlinear time/chemical strength/exposure function from the chemical development of the negative. Then to get from negative to print, black and white photos undergo another pair of multiparameter nonlinear functions (representing the exposure and chemical development process), possibly including intentional manipulation by a human operator. Depending on the film, filters, development process, and printing techniques used, this can be relatively close or quite far from a naively converted 3-channel color picture. For details cf. for example Ansel Adams’s books The Negative and The Print.