>At 12:42 PM -0400 6/1/06, Manfredi, Albert E wrote: > >I think you're confusing two different topics: whether entropy >necessarily is reduced as resolution is increased, and how much >information a human can absorb in an image. Not really. I agree that the amount of entropy in a video stream is variable. In the early phases of the ATSC work there were no 720P cameras available for testing. Europe had developed a 750/50P camera ( I believe it was 750 lines). Anyway, this camera was so noisy that it was very difficult to determine anything meaningful about the ability to compress this format. It was not until late in the process that a decent 720P camera surfaced, which showed the viability of the format. If an imaging system introduces entropy, the level of compression efficiency will suffer. But this is different than than the theoretical case i was discussing. If the level of entropy is held constant as the resolution increases, the compression efficiency will improve and virtually all types of source material. Remember, most of what we see is at relatively low frequencies - in most areas of the image the sample accuracy will improve with improved resolution, without adding more information (detail) to compress. I would also note that compression efficiency typically improves with frame rate, although there are more frames to compress. The work that we did with 72P demonstrated that there is little additional overhead for this frame rate compared with 60P. The reason is that there is less temporal blurring and and more accurate samples for use in the motion prediction routines. As for the human aspect of this equation, it should be obvious that higher resolution source can be compressed more than low resolution source. As you increase resolution you more precisely localize the information in the coding blocks - i.e. there is less detail in a block, thus it can be quantized more without having a significant impact on the perceived image quality. Think of it this way. In a cell phone image, If I quantize an 8 x 8 block to a single value i will see a very visible 8x8 block in the image (assume 240 samples per line, so the block represents 1/30th of the line). In a 1920 x 1080 image, that 8 x8 block represents 1/240th of the line, thus it is only likely that you would even notice the quantization artifact on a VERY large screen. Case in point. Back in the late '90s at a SMPTE conference in Pasadena, we watched a Shuttle launch live in 1080i HD. Unfortunately the large auditorium projector was not set up in time for the live event, so we watched it on a variety of displays, none of which were large than about 60 inch diagonal. The images looked spectacular on all of these displays. That evening they got the theater working and we watched the launch again on a big screen (about 10 foot diagonal). This time the blocking artifacts were readily apparent as they were no longer being masked by the fact that they were suppressed - relative to a human observer - since the pixels were much smaller at the nominal viewing distance for the smaller screens. All this points out is that it is highly desirable to have VERY ACCURATE samples for compression, if you want to deliver the best image quality, AND it is highly UNDESIRABLE to quantize the samples to the point of changing them such that the human observer can detect artifacts. The best way to achieve this is oversampling during acquisition and production, then resampling of the final product to a lower resolution for emission. We are JUST getting to the point where it is possible to produce very good quality 720 images for compression. We can do a pretty decent job with 1080@24P when it is captured on film then re-sampled with a telecine. We are still years away from having good 1080@30i cameras, and even further away from 1080@60i. To do these "emission" formats properly we will need to sample at at least 3K x 1.5K (1.5 X oversampling) - 4K x 2K (2 X oversampling) would be even better. >This: > >> Increasing resolution typically allows for much >> higher levels of compression efficiency due to the >> improved correlation of the samples, and the ability >> to restrict the content in a coding block to samples >> that are highly correlated. > >assumes that you don't need to accurately portray all the added >information the higher resolution image COULD impart, because presumably >a person can't absorb it all. There is a great deal of truth in what you say. I would only add that if you quantize too much you will introduce artifacts that will be visible. So the key is to have enough headroom in the emission channel so that you never stress the encoder. > >As others have said, given a large screen and close viewing distance, >*and* the fact that you don't know what spot of the screen the viewers >will focus on at any given time, it's hard to make the case that the >higher res image can necessartily be compressed more "efficiently." Not true. I agree that it is important to run the compression algorithm such that we never produce visible artifacts. But the physics are very straight forward here. If you have two cameras framed to capture the exact same image, the camera with higher resolution will produce an image that can be compressed more relative to the lower resolution camera all things being equal. The localization of detail in the coding block alone is a big benefit, but motion compensated prediction routines ALSO work better because of the localization AND the improved accuracy of the samples. We are not examining still images here, these are moving images with a temporal update component. This component helps to average out the perception of artifacts. The key is to keep the level of quantization noise low enough that it looks like the typical - slightly noisy - video we all grew up with. Also note that the Hollywood types like film grain, which is perceived much the same as noise. > >If I send an image of a tree to a cell phone, the leaves are allowed to >look like a green blur. If I send that same image to a UHD display, I >want to see individual leaves, the veins in the leaves, the tiny >droplets of dew, and a detailed rendition of the insects chewing on them >as well. So really, it's not at all clear that compressing this image >has become any easier. The closer you look, the more detail there is to >encode. I don't want to see a green blur on the cell phone. I want to see some texture in the tree, which will be there if the image is properly downsampled prior to compression and NOT overly quantized. Even with UHD you are not going to see all of the stuff you describe unless we zoom into the tree. A better analogy would be to look at one leaf. On the cell phone the dew, bugs and veins would have little detail, but we would see them. On the UHD display you would see the tiny hairs on the surface of the leaf, and be able to eye-track the bug as it walks across the leaf. There IS NOT infinite resolution in nature, although one could argue that with enough resolution we could actually see the electron cloud surrounding the nucleus of an atom. Most of what we see is lower frequencies. When we want to see more detail we move closer, and in the case of many on this list like myself, we put on our reading glasses. I don't wear reading glasses when I watch TV. I let the camera zoom in for close-up views, and i do not need resolution that equals reality in order to be entertained. The Hollywood types fully understand this and expend great effort to limit the resolution in their products. > >Obviously, if your camera has a poor lens, and the added detail cannot >be rendered, compression ratio will increase. This may be a factor, but it is not obvious. If the level of blurring is properly matched to the format, this should not be the case. Unfortunately, cameras with poor lenses typically are limited in other ways as well. We do not expect an HDV camcorder to equal A Sony Cine-Alta camera. The HDV source may be MORE difficult to compress at the same bit rate due to noise and other sampling errors. > >On the multicasts making DTT more attractive, I understand that you >might not think that rerun of old shows is not compelling (to you). I >didn't mean to imply that's all one can transmit on the multicast. It's >still clear that well-selected multicasts will make DTT more appealing >than just providing what's available with NTSC. Chairman Martin was not >coaching his remarks based on a specific choice of multicast content. He >simply said that multicasts would make DTT more attractive. It all depends on the content Bert. Boradcasters still live in the land of surfers and browsers, who just turn on the TV and click through the channels until they find something interesting. In a world where there is limited choice, they will always find an audience for re-runs. But this audience will be small compared to fresh, highly desirable content. Many of the cable channels are browser channels disguised as vertical niche channels - e.g the Sci Fi channel is mostly old Sci Fi re-runs. This works because they have a national audience available to upwards of 70 million homes. If broadcasters were to compete with cablem rather than relying upon it to reach their viewers, I would agree that multicasting makes DTT more attractive. Unfortunately, at the moment DTT is more like a wallflower sitting through the dance while the beautiful people party. Regards Craig ---------------------------------------------------------------------- You can UNSUBSCRIBE from the OpenDTV list in two ways: - Using the UNSUBSCRIBE command in your user configuration settings at FreeLists.org - By sending a message to: opendtv-request@xxxxxxxxxxxxx with the word unsubscribe in the subject line.