||Thread Tools||Search this Thread|
|Oct 3, 2007, 12:15 PM||#11|
Join Date: Jun 2006
Video footage analysis.
Pleasant Surprises, and some disappointments:
What I am looking at:
I have completed an analysis of the footage, it is based on good consumer quality expectations and low end professional (broadly similar). This is an lot less than the standard of an truly top end professional review. If you can notice it on an super TV, it should not be there, it should be unmarred, even if there is loss of contrast in detail, and some loss of detail, as long as it is casually unnoticeable. This gives preference to the accuracy of shape of form and left over detail. Fast moving detail maybe reduced, if unnoticeable. Low end consumer is an lot less than this. In truly professional quality, the codec would play with the more unnoticeable parts of the image (visually lossless to true lossless). The data-rate for truly professional level might be double the level for good consumer quality in the same codec (this is also some overhead to allow for processing the video).
Problems with accuracy of the images being analysed:
Unfortunately, because of the trans-coding to HDV codec, it is hard to tell what is caused by the original codec, and what is caused by the trans-coding to HDV (High Definition XDCAM is probably better suited, or something that is Intra visually lossless codec, in any case "no loss" is not guaranteed).
There is also the problem of the standing machine I am using is low powered, and the player does not stop where I want it to, to take the still, and perhaps it is stopping at the GOP border frame often (the best rendered frame that the rest of the frames are based off of). The in between frames use difference compression (recording the difference from the main frame in the group) and can look an lot worse. So the first frame is not representative of the lower quality of all the frames in between. The variance between the frames I have been able to detect, might be that the best frames look double the quality of the worse, but this might vary an lot more, depending on the scenes compressibility under difference compression of the in between frames.
So, an lot of qualifying "if", "but"s, and assumptions, will be used, but if anybody can verify the truth of these things on their better equipment, please do.
You should refer to the Japanese article for further information on these subjects, and also on still and camera control features. This post is purely on analysis of the video produced in the article.
Sensor Noise, Noise reduction, Black clipping:
Significant sensor noise still looks like it is still there, though noise removal (or is it also the trans-coding) is pretty good at reducing it to an light crawling on mid to darkish scenes, and primary colored objects (purer red, green, and blue, that causes the other colors to be dark, containing noise). It seems to be there in the stills, and seems to be at an level similar to what has been seen on the previouse HD cameras (except that the noise removal has reduced it in the video). it is not too bad (but not really professional). Blacks in well lit scenes seems to be more clean, is this clipping or noise reduction at work?
The sort of leftover crawling pixels caused by noise removal, still requires encoding space that takes away space from real details. Low end codecs have an lot less data space to devote to an image frame, noise is generally irregular, and so can't be compressed well, if it is compressible at all. So rendering it in the scene sucks up huge amounts of data space that is no longer available to be used to render the authentic image. Where it gets too dark, cameras start adding gain to boost the brightness, this increases noise even more, sucking up more data. Even though noise reduction is reducing this amount, consumer sensor manufacturers have to learn that the only way to really go, is to have sensors that have virtually no noise at any level with 0db gain, to start with (professional cameras can have many times less noise again at higher gain). This produces an cleaner image to start with.
It looks good compared to the previouse cameras, probably acceptable in consumer or professional terms (though there is no bright tropical midday midsummer sun scenes yet to test this). You can get both well lit open places and under the tress exposed at the same time. Whoever, such as with the blown out flowers in the scene with the girl, it may not be enough and needs to be exposed for.
Too much for bright colors, they seem to glare out. I do believe there were modes that were less vivid on the HD1 series, and hope they are on this one as well.
However, on the sample frames given in the article, the telephoto frames seem washed out, I have not examine them, and don't know why.
Artifacting, Blocking, Fly screening, detail shape preservation:
Blocking during panning, looks very low, and blocking because of detail is perhaps much less than half that of the HD1 series, and appears to be acceptable for consumer application. Much detail seems to be preserved compared to the HD1 series. However, given more than an little movement in frame (panning is better) blocking and blobbing appears to be there, and water shapes seem to have squarish shapes and blocks. It looks much better than the HD1 series overall, but this needs further testing.
"Blobbing" is an term I am giving to the feature of the H264 codec that attempts to de-block an image by blurring out the block, producing an ghostly looking blob.
Fly Screening, looks similar to looking through and Fly screen and moving around, causing pixels to flash and details to staircase. Cannot tell, seems to be there, but could be masked, or produced, by my equipment or trans-coding.
The images have silky smooth tones, and good colorization (though overly saturated) producing the beauty of the images.
Can't really tell from the clips, but noise is abundant and not completely dealt with. The colorization and tonality still look significant. However, compression/movement issues look more because of the noise (which is fairly normal).
There seems to be significant flare (is that the proper term?) chromatic aberration on sides of objects that is away from the center of the lens, and that gets worse in all directions with brightness. Tends to be bluish on the outside of objects and reddish on the side closer to the centre. It also results in strangely shaped looking blur around objects. This could look disconcerting. I do not know if there is any other problems with aperture shape (not examined) producing this look.
Can't say, was hoping ti was consumer Altasens, but from the altitude and noise levels, I would think not. But then again, I do not know the performance of the new consumer Alta-chrome range of sensors. So, details are unknown. It is 4MP, so one of the modes is interpolated, which affects resolution. The noise is one area of improvement.
Codec looks consumer level (but HDV trans-coding of the clips getting into the way again). It needs more codec data rate, and an better implementation of the codec that serves shape preservation of form and details. In this day and age, Blu Ray and HD DVD offer h264/WMP at much higher bandwidths than the present h264 cameras out there, and it could easily require an 50mb/s h264 Inter camera to match the processed images those processed disk formats can provide. So, 12mb/s is not where we should expect to stop with an consumer camera in future. The sweet spot, I think, is between 18-36mb/s.
Beautiful artistic looking footage. The image produced is delicate, from the tonality delivered, and beautiful, despite the over saturation.
What we wish the HD1 had been.
Consumer level, even more consumer under codec stress (it is the soccer/sports filming I am worried about). But great image colorization tonality, and latitude possibilities. Image like an good consumer camera, but compression is less desirable. This camera is still an big step in the right direction for Sanyo.
In detail, examination of the clips:
The main trans-coded clip:
First scene, Shoot moving from ground up to Girls face.
Blocking level on the dirt is good, retains an lot of detail. Some detail breaks down to blurry micro blobs (an "feature" of the H264 codec. Did see some blocks on lower neck/upper chest. You will remember that HD1 series had much much more problems. What is very good, is that this is an movement scene, and it did not break down to large amounts of blocks as new details was revealed, something I think would produce large amounts of blocks on the HD1 series. However this is not on complex, or large complex movement, but still an big improvement. I am assuming the trans-code to HDV did not do de-block filtering.
Hardly any crawling pixel problems on plain surfaces as experienced in HD1 series.
Latitude burnout on Girls left Arm (Burnout defined as level that is so bright as to affect detail). But quiet acceptable overall for an consumer camera, even low end professional (assuming that trans-coding has not caused it).
Bright colors are too saturated, some scenes it is alright, others, such as the girls striped coat, and grass, it looks too much. Pushing my monitor from 73 to 100 brightness made the saturation look more natural. Color balance, I can't really judge too much.
In the out of focus background, we can notice something strange about the shape of blur around objects. This blur seems to be misshapen. On closer inspection, it seems to be exaggerated on the side away from the center of the lens, into flary chromatic aberration. We can only hope this is temporary and production cameras will be free from it. It varies depending on intensity of the brightness of the objects producing the flare, until it goes into an big blue hallo. This can be seen in the scene of the tree and leaves, where the blue sky shows through the leaves. On duller objects, such as the flowers behind the girl (that are also disconcertingly burnt out) it is bluish tingled on the outer flared side, and reddish on the other. Altogether, disconcertingly irregular.
The butterfly scene:
Beautiful, but notice that there is an bluish noise tinge throughout the out of focus background, but still beautifully tolerable. In many scenes you will see this sort of slight crawl on mid to darkish levels.
The leaves with bright yellow leaf in the middle:
The leaf is burned out so much that it is nearly completely burned out, and detail is reduced.
But, in this closeup in focus scene, some slight fly screen type crawling (aliasing and flashing) of edges of details can be seen (but am not certain). I don't know if this is my computer causing this when displaying 1080i on an window an quarter the size, or is caused by the trans-coding (that actually should even improve fly screening because HDV is downscaled 1080i). But I would not expect fly screening at all in an downscaled 2MP image from an 4MP sensor. So, I don't really know if it is less or more in the original. Maybe somebody else can answer (I can't playback original h264 1080i clips on this computer).
What is fly screening, it comes from technical problems with many sensors. Sensors are made from an square grid of light sensing pads, that are red, green or blue normally, and an image is formed and interpolated from what is projected across this grid by the lens, as many of us already know. The problem arises from two sources. One the pads don't touch each other, there are gaps between them, and they might not even be truly square, as some circuitry is between. This is an bigger problem on CMOS than good CCD's, though advanced CCD's also have lots of circuitry as well. When the border of an detail or edge, crosses in between the pads, it cannot be seen, and the pixel it left goes darker, and when it passes to the next pad it goes brighter (causing the flashing of the pixels) also it can cause crawling aliasing/stair-casing along edges and other things. The other main problem is the way red green and blue pads make an image, when an object that is closer to one of these primary colors is displayed, then the other colors receive less light and act an bit like an gap, producing more flashing. In an still image it is not that noticeable compared to an moving image. The amount the pads fill the grid, is called the Fill factor. There is one CMOS company that has an technology that purportedly fills in the gap between the pads, called Fill-factory, now owned by Cypress Semiconductor. But in general, all cmos and good CCD's suffer this. To reduce this they use micro-lens arrays across the pad grid, to focus light away from the gaps towards the pads, but in general, even if the sensor has this it usually doesn't get fill factor near 100%. The downside is that the faster the aperture, the more it hits the lens wrong and causes aberrations and washout (but using the Fill-factory technology instead allows super fast lens to be used).
The Sitting Cat scene:
Around the movement of the cat licking, can be seen much blocking and blurring. This might be expected, because of the complexity of details in the cats fur diverting data-rate to render it (yes it can be that delicate of an balance). Compared to the still of the same scene, there seems to be less detail, or something, but we see that there is increased noise.
Notice the large patch on the concrete edge below and to the left of the cat, complete burnout again, but this is an high latitude situation. It serves the consumer situation well enough, but I wonder how it would go if ti was the stronger light of an bright tropical mid-summer, mid-day, sun coming through the trees.
The concrete is blurry in patches, hiding blocks again, but this is much better compared to what I have seen on the HD1 series. It is tolerable, I should say, between cat and concrete, dirt at girls feet and Girl, even the next duck scene to some extent. I would guess it is less than half the blocking of the HD1 series at least (that includes the blurred out bits). Except for water, where there is this high movement and it breaks down, it probably is not so noticeable, except that it looks irregular detail.
Scene, Water reflection of umbrella:
Much blocking blurred out by codec (I hope this blurring is marked in file, to happen at playback, because post tools could do much better de-blocking/detail restoration, algorithms.
Even here, the blocking once again, seems to be an lot less than you would get on the HD1 series, and an idea of the reflected detail of the scene is there, however it is not good enough, but adequate. But once again, how much is the original and how much is the HDV encoding.
Duck on Water scene:
The water seems to shape well. Squared shapes in water movement, and blocks in moderate movement areas. Outside the area of main turbulence, there are an few blurry bits as the codec tries to blur out blocks.
On background, well lit, still water, there is light crawl of blocks again, possibly from noise that has been removed.
The low light HDV sample:
The background still crawls an lot after noise removal. The magnetic ball device shows blocking and movement compression problems, even though it is the only thing moving in the scene, and is an small part of it. This varies in between frames, and catching the main ball entering the frame it is worse. However, I would have expected an better rendering on the device, it may well be caused by the large crawling background.
Why does this matter, as problems in that sort of movement is not so noticeable? If it can't do this under ideal conditions (for lowlight) it is probably not going to be any better in normal scenes containing lots of movement.
The native footage water torture samples:
I don't know how much is there and how much is the VLC rendering playing up, but from looking objectively, and with other players, I think that there is some underlying disjointed blocking in the water, and harsh looking shapes.
That scene is definitely consumer quality, in no way professional. The consumer quality here, is not what I would call high, and the camera has achieved much better elsewhere. Much of the movement of the water shapes have an more predictable flow here, it is not the ocean surf, or the surface of an choppy lake (and the lake scene at the end of the main HDV encoded clip does not count, as it is not very choppy).
Can't really tell noise on this clip, as the water is over everything moving the image, except for the grey/white parts that should exhibit the least amount of effects of noise. The noise looks less than on the HDV encoded clips, so who knows (and I am assuming that they didn't have gain/exposure turned up on the other clips, that produces more noise). We've come an long way baby.
|Thread Tools||Search this Thread|