Sunday, September 18, 2005

Learning about the Canon XL-H1

This camera caught me by surprise as many of the camera vendors now contact CineForm before they announce their HDV camera (even many non-HDV vendors discussion CineForm implementations.) However this is a very nice surprise, and we all expect Canon to have a high quality solution. As the XL-H1 is a 1080i HDV camera it will be closely compared with the Sony HVR-Z1; neither camera offers a progressive imaging sensor, yet both have a pseudo 24 frames implementation -- speculation on this is great as the Sony method under-whelmed the market, Canon may/should take a different approach (see my speculation below). The good news is this new camera should operate flawlessly with all CineForm products as it complies with existing HDV and HDSDI (bonus) standards.

The clear difference between this unit and other HDV cameras is the addition of HDSDI. We already have several customers using live HDSDI feeds into Prospect HD workstations for onset capture, bypassing the camera compression of HDCAM/DVCPRO-HD etc, so the same work-flow will apply for the Canon XL-H1. So just like the high end HD solutions, the HDSDI feed is pre-compression, the raw data from this HDV camera is the highest quality feed, perfect for Prospect HD's 10-bit CineForm Intermediate compression (the HDV tape acts as a backup only.) There has been speculation whether the HDSDI feed is 8-bit or 10-bit YUV 4:2:2. I believe it is more likely to be 10-bit, counter to a few opinions, as there is no technical reason not to output 10-bit. The RGB CCDs capture light in a linear manner with 12 to 14 bits of precision (depending on the sensor), after in camera processing the resulting RGB linear data is converted to YUV 4:2:2 in a 709 colorspace. The RGB precision is plenty to support a 10bit YUV gamma corrected output. The top 8-bits would then be sent on to the MPEG compressor (where it is further down sampled to 4:2:0.) If the processed linear RGB data is converted directly to 8-bit YUV, that would have been an error on the part of Canon engineering (so I don't think they would have done this), particular as HDSDI is inherently 10-bit (8-bit data is sent as 10bit with the two least significant pixels set to zero.) So as soon as cameras are available it is straight forward to determine the bit-depth of the HDSDI feed.

As we will get an uncompressed 1920x1080 out from HDSDI (higher than HDV's 1440x1080) the sensor resolution is significant. Canon has increased the resolution over the Sony Z1's three CCDs of 960x1080 (which use a horizontal pixel shift of the green sensor to increase the effective resolution to approach that of 1440.) Instead the XL H1 has gone with 3 16x9 CCDs of 1.67M pixels each, suggesting a sensor resolution of around 1720x970 (assuming square pixels.) I just read that the effective sensor area for HD is 15.6M pixels, which would indicate of non-square pixel sensor of 1440x1080 (which makes better sense.) If the lens allows, this should resolve a sharper (or detailed) image. The down size of so many picture elements on a 1/3" CCD, is a reduced light sensitivity. Sony the 960x1080 CCD, with their extra elongated pixels, have a larger area to collect light, and therefore potentially slightly more sensitive.

The most anticipated element of the XL H1 will be how it achieves 24 frame acquisition using a interlaced sensor. Canon has been up front with the 24 mode not being acquired in a true progressive sense, but in one statement it seems it will not have the motion jutter of Sony's CineFrame 24, which suggests the sensor may be run at 48Hz rather then 60Hz for the 24F mode. This will be an excellent first step. However if all data is interpolated from a top or bottom field, the vertical resolution can't exceed 540. I don't give much weight to an intelligent de-interlacing process, as that will potentially introduce motion artifacts as the good algorithms would take too long for the camera to compute. Canon could generate a good resolution pseudo progressive image if they ran the green CCD with reversed field dominance (output a bottom field when red and blue CCD output a top field) this way the same pixel shifting technology that makes Sony's 960 res CCD output near 1440, could help the three 540 fields (red-blue and pixel shifted green) could achieve a progressive vertical resolution approaching 810 (which is about the maximum a normal interlace picture can achieve any way.)

One last note on the 24F mode. The "Advance" 2:3:3:2 pull-down of 24p DV is unnecessary in HDV as there in no frame compression used in the 1440x1080 mode (spec HD2.) As only field compression is used, the standard 2:3:2:3 pull-down will work perfectly well. A true progressive frame signal can be exacted with ease.

P.S. 10/12/05 -- Now that we have a camera I have posted some real world data.

Saturday, September 17, 2005

So many new HD cameras to choose from, so much misinformation.

There is much heated discussion on the various HD and HDV forums regarding sensor resolution, compressed types and bit-rates of the new crop of prosumer HD cameras. So much is focused on these elements, yet the numeric data is often less significant than the quality of the lens and the physical size (not resolution) of the CCD/CMOS. This is encouraged as so much of the marketing of these cameras is rooted in the numbers -- three CCDs is assumed to better than one, chroma sampling 4:2:2 better then 4:2:0, and 100Mb/s is better than 25Mb/s, all seems so true, yet in most cases the impact of image quality is far more complex than that, with the numbers often hiding more significant characteristics.

What is the true resolution of a camera? Put a resolution chart in front of its lens and you have better information than any of the marketing data. However the ability to resolve detail will be impacted by the sensor resolution, physical pixel size, compression resolution and chroma sampling. I in the next few post I intended to go into interesting technical details on a range of these new cameras, pointing out where I think the marketing information based on the numbers are either misleading or significant to the final out image quality.

Here are the range of the sub-$20k HD cameras that are annouced or currently available.

Canon XL-H1
Grass Valley Infinity
Panasonic HVX-200
Sony HVR-A1
Sony HVR-Z1

I intend to discuss a little on each. Please point out if I have missed any interesting new cameras.

Friday, September 09, 2005

Extending the Prospect 2K work-flow on Lbs

Our next major film project Lbs will be using the upcoming Prospect 2K editing engine. Prospect 2K is a work in progress and we aren't currently marketting or selling it. I will occassional discuss the product's updates as will learn from the post-production of this film.

One of the post-production requirements is to monitor the timeline via HDSDI even though the scanned frame resolution is 2048 x 1276 , larger frame size than HDSDI generally supports (at full frame rate.) As the feed is primarily used for color correction we will window the 2K source so that most of the image (80%) can be seen across HDSDI. For this project will are sending out the center 1920x1080 image. This HDSDI feed will pass though a BlackMagic HDLink box to enable the mapping of log density image for accurate viewing on an HP 24" DVI LCD panel.

All of this has been working perfectly well. The trick now is to come up with the necessary curves between the log images and the 2.2 gamma LCD such that color correction performed, with this simple workstation setup, will match the final film out. This will involve correcting for the errors / distortions in the film final out -- we will be performing film out tests and using that data to help generate the curves. This is the next big stage of this project, and no small feat. It should be a lot of fun.