What's an example where you wouldn't want to clamp point processed values?
rbustama
@hiliang HDR panels were briefly mentioned in lecture, and I think they can support color values larger than 8 bits. I think not clamping values (to 255) would allow the entire range of an image to be expressed on an HDR panel? Not sure, but could be one example
hiliang
Later on in this lecture, preserving negative values for additional convolutions was mentioned, I think an example was allowing for intermediate negative weights to carry over.
I'd assume 8 or 10 bits would be a formatting limitation of the final output?
motoole2
When working with images, we would typically (1) read in an N-bit image as input and convert it to floating point, (2) do whatever computations are necessary, then (3) quantize the image using an M-bit image to save/visualize the result. Clipping is often only done at this final step, before visualizing results or saving images.
Although N and M are usually 8 bits, this is not necessarily the case (as pointed out by @rbustama). Your images might be of higher bit depth, or your display may support higher bit depth. Btw, one common HDR (high-dynamic range) format is .exr, which uses 16- or 32-bits color depths.
What's an example where you wouldn't want to clamp point processed values?
@hiliang HDR panels were briefly mentioned in lecture, and I think they can support color values larger than 8 bits. I think not clamping values (to 255) would allow the entire range of an image to be expressed on an HDR panel? Not sure, but could be one example
Later on in this lecture, preserving negative values for additional convolutions was mentioned, I think an example was allowing for intermediate negative weights to carry over.
I'd assume 8 or 10 bits would be a formatting limitation of the final output?
When working with images, we would typically (1) read in an N-bit image as input and convert it to floating point, (2) do whatever computations are necessary, then (3) quantize the image using an M-bit image to save/visualize the result. Clipping is often only done at this final step, before visualizing results or saving images.
Although N and M are usually 8 bits, this is not necessarily the case (as pointed out by @rbustama). Your images might be of higher bit depth, or your display may support higher bit depth. Btw, one common HDR (high-dynamic range) format is .exr, which uses 16- or 32-bits color depths.