Dialogue Concerning Two Imaging Systems

Dialogue Concerning Two Imaging Systems Galileo Galilei November 24, 2009 Sagredo: HDR is definitely a cool technique, i just don’t get why they all look so HDR-y (instead of just looking correctly exposed everywhere) Salviati: @ Sagredo: good observation. This is probably because the wide dynamic range captured with the technique must ultimately be compressed to a much narrower range for storage and display. (E.g. 8bpp JPEG.) Since this is done differently for different parts of the image, the end result looks unrealistic. The best thing to do would be to somehow capture all of the dynamic range in the original image in one exposure and present it on some kind of projectable transparency. #grandmaglance–that’s basically what slide film does. Salviati: Also, HDR needs to get beyond its "look at me, I’m HDR" aesthetic. It’s really just one more tool in the ever-growing digital photography toolbox. Sagredo: "This is probably because the wide dynamic range captured with the technique must ultimately be compressed to a much narrower range for storage and display." – i don’t get this concept take [Figure1] and [Figure 2] for example. in the first correctly exposes the sky and underexposes the sign, and the second overexposes the sky and correctly exposes the sign. intuitively it seems like an HDR image could take the sky from the first and the sign from the second without compressing anything into a narrower range or whatever Salviati: "Intuitively it seems like and HDR image could take the sky from the first and the sign from the second." That’s what it does. And that’s why it looks unnatural! We’re probably pretty good at understanding light and shade in visual perception, so we can quickly recognize an artificial scene. This is probably why photographs taken in lighting conditions remote from our experience look fake, e.g. those taken on the Moon. Sagredo: but i don’t understand why JPEG would have any problem storing this kind of dynamic range Salviati: Because it is 255 (bits) per R, G, and B. That’s only not nearly enough range to realistically represent a scene where the darkest point might be many stops dimmer than the brightest point without compressing it "into a narrow range or whatever." A good

Figure 1: Tom’s image correctly exposing the sign.

Figure 2: Tom’s image correctly exposing the building.

dialogue concerning two imaging systems

digital camera sensor, for example, should capture 12 or more bits per channel. Sagredo: so what’s going on [in Figure 3]. (i copied the sign from the image in which it was correctly exposed) Sagredo: PUT ANOTHER WAY: i can increase the exposure on dark areas without blowing out light areas in lightroom by moving the "fill light" slider right. however, this introduces artifacts when you move it too far right. isn’t HDR just extreme fill light without the artifacts? (even if it isn’t, couldn’t it be? i.e., if the JPEG dynamic range supports fill light, why can’t it support the full dynamic range of an HDR?) Salviati: I don’t have aperture, so I’m going to have to guess what the "fill light" slider does. It sounds like it is selectively brightening the dark areas. HDR tools selectively merge bright and dark areas from different images, so the results can be similar. The HDR look doesn’t come from this process but from the "tone mapping" necessary to make the merged file viewable on displays with limited dynamic range.. This is the process I alluded to above. (http://en.wikipedia.org/wiki/Tone_mapping) Using your sign example, suppose the real scene has light values on some arbitrary scale of "EV," where each EV differs by one aperture stop. (Remember each stop represents 2X more light than the stop before.) Perhaps the sign has values from 16 to 20 EV and the darkest part of the scene has values from 1 to 4 EV. You capture correctly exposed images of each so you now have two digital files. Assume for simplicity’s sake that your digital files have a range of values from 0-255. A value of "255" in the first image might correspond to 20 EV of real light in the first exposure, while it might mean 4 EV of real light in the second exposure. That means first image represents real light amounts 65,536x up to greater than the second. The tone mapping algorithm has to merge these values that represent radically different real light ranges into the same image. It could do this by mapping the first (bright) image onto a range like 250-255 and the second on a range from say 1-2. But that would be unsatisfactory because those values would appear only about 250x brighter and lots of tonal information would have to be discarded (255 values would have to be sampled down to 1-2 values). Instead, tone mapping algorithms typically combine the values into one image (so that 255 might represent 20 EV in one part and 4 EV in another part) and then use perceptual tricks to suggest great contrast between the values. For example, to make a building photographed against a bright sky appear darker than the sky, the algo-

Figure 3: Tom’s image correctly exposing the building.

2

dialogue concerning two imaging systems

rithm might brighten the sky immediately surrounding the building, creating local contrast. Those tricks are what make the final file have such a strange appearance. Personally, I think most images produced with perceptual tone mapping are pretty ugly. And the required multiple exposures rule out street photography, wildlife photography, and anything else with potentially moving subjects. You might ask why we tone map down to 255 values in my example when there are file formats that support many more bits per channel. That’s correct, but you will ultimately need to collapse your image down to approximately this range for printing or display. (Unless you are using transparency film or you have a high dynamic range monitor. See http://www.bit-tech.net/hardware/2005/10/03/brightside_hdr_edr/1)

3