Test Drive: Affordable HD Formats, Part 2

 Continue Reading

Test Drive: Affordable HD Formats, Part 1

A funny thing happened on the way to the second installment of the first edition of the newly renamed Affordable HD column (see part one). I reconsidered and changed plans. That is, in the original edition, I explained that because I had a host of affordable HD cameras in my grimy little hands—including HDV, AVCHD, and DVCPRO HD—I would compare the formats, a process I started in the first installment when I identified the technical specs of each format and discussed their strengths and weaknesses.

After a little thought, however, I recognized that I had to decouple the camera from the format to assess the format, not the format/camera combination. For example, if I shot the same resolution chart with four different cameras and then compared the results, there would be two components to the quality equation: the camera and the format.

Looking at it another way, while I love my Canon XH A1, it''s getting long in the tooth by camera standards, and it might be unfair to render judgment on the HDV format on the basis of that camera. What to do, I wondered, my deadline rapidly approaching (past, actually). How to test the format exclusive of the camera?

Table 1: Formats and features.
Click here for a larger image

I glanced back at the relevant specs for each format and realized that there were two components to each spec: the maximum capture resolution and the codec and its respective parameters. For example, just looking at the numbers, you would assume that full-resolution AVCHD could capture and retain better detail than either other format because it stored the video in its native resolution, while the other formats subsample the file horizontally then zoom it back up for display. On the other hand, you''d expect DVCPRO HD to show fewer (if any) compression artifacts because it''s produced at up to four times the data rate of the other formats and uses I-frame-only DCT rather than long-GOP formats as HDV and AVCHD do.

Figure 1. Comparing frames from the encoded files with a frame from the original.
Click here for a larger image

Of course, I don''t get paid to postulate, I get paid to test and prove or disprove my theories. So here''s what I did. I dusted off my Canon Digital Rebel SLR camera, which shoots an 8-megapixel image. I then shot my standard resolution chart, DSC Labs' ChromaDuMonde, in raw mode. Then I imported the image into Adobe Photoshop, cropped and scaled the image down to 1920x1080, imported the frame into Adobe Premiere Pro, and produced a 10-second video, saving it using a lossless codec.

Next, I loaded the file into Rhozet Carbon Coder and rendered the file out into four different targets: AVCHD at 1440x1080i (14Mbps) and 1920x1080i (24Mbps); HDV at 1080i (25Mbps); and DVCPRO HD at 1280x1080 (100Mbps). Then I grabbed screens from each file and compared the quality (see Figure 1).

At the risk of sounding hyperbolic, I was shocked at how little difference I saw between the formats. Essentially, DVCPRO HD discarded 33 percent of the horizontal pixels during encoding, and then it was able to reconstitute the frames during decompression with just a hair less detail than full-resolution AVCHD. You''ll have to view the full-resolution image to see and believe it, but the difference between any of the formats was virtually imperceptible.

So scratch storage resolution as a source of format differentiation, and let''s turn to comparing the effectiveness of the compression technique used by the respective formats. Here, while the single frame of the ChromaDuMonde chart performed admirably for the resolution tests, I needed a source for the compression-related tests.

Figure 2. What the heck is a screenshot of RedCine doing in an edition of Affordable HD?
Click here for a larger image

Obviously, I couldn''t use any of the cameras as source, since that might bias the tests. In a perfect world, I would just grab my Red Digital Cinema camera and go shoot some hard-to-compress footage in 4K mode, then encode that footage into DVCPRO HD, AVCHD, and HDV. Unfortunately, as our shrinking 401(k) balances keep reminding us, the world isn''t perfect and my Red camera hasn''t shown up yet.

So I did the next best thing, and surfed over to www.redrelay.net and downloaded some footage shot by the Red One camera. This footage is stored in formats ranging from H.264 to R3D, which is the raw Red format. Given that my deadline was way back in the rear view mirror, I was a bit loath to attempt to unlock the secret of the uncompressed R3D format, which would give me the cleanest starting point. Still, I dived in, and I was rewarded by a surprisingly painless workflow.

That is, I downloaded a beta version of Red''s own RedCine software. It converted the R3D file into a QuickTime file with several format options, including the Cineform codec, which is available as a free 15-day download—more than enough for my needs. I exported at 4K resolution, which input quite nicely into Premiere Pro so I could edit and output as desired. Problem 1, obtaining my test footage, was solved.

Now I faced problem 2, which was encoding the files in as close to an apples-to-apples comparison as possible. Specifically, I had multiple encoding tools at my disposal—lincluding Rhozet Carbon Coder, Adobe Premiere Pro CS3, and MainConcept''s reference encoder. All could produce the required formats, but there was always some niggling detail that made me uncomfortable—initially relating to DVCPRO HD.

Specifically, all of the encoders treated DVCPRO HD like the black box that it is—with very few, if any, configurable items. The MainConcept encoder could only produce at 1920x1080 resolution, which wouldn''t work because in practice DVCPRO HD is typically used at 1280x1080. And Premiere Pro and Carbon Coder could only produce interlaced output, which really stank because the Red footage was all progressive. Obviously, I couldn''t produce DVCPRO HD in interlaced mode and other formats in progressive, so I ran the first experiments in interlaced mode.

Obviously, interlacing wasn''t an issue with the resolution chart because there was no motion, so two fields merged together into a perfect frame. With high-motion video, however—which I was seeking for my compression test clips—interlacing would produce noticeable aliasing in the video, which would be distracting and might obscure other artifacts.

Figure 3. Comparing the formats in interlaced mode.
Click here for a larger image

As before, I used presets in Carbon Coder to produce the four files in the three formats, then loaded them into Premiere Pro for snapshots. After testing about four different clips, the nickel dropped, and I reached the rather obvious conclusion that all three formats at the format specific data rates were very artifact-resistant; otherwise, the vendors wouldn''t have used them.

To put this quality in a bit of perspective, at 24Mbps and 1920x1080 resolution, AVCHD devotes about .385 bits of data per pixel. By comparison, a 720p streaming file encoded with the same H.264 codec might have a data rate of 2Mbps, which is .0723 bits per pixel. This is a long way of saying that not only are the compression technologies used in these formats very advanced, the data rates are very generous compared to other common uses. That''s why these cameras produce such stunning (and affordable) HD quality.

The clip that showed the most obvious quality difference was an underwater clip shot by Kenneth Corben. Here, after zooming each clip to 300 percent and comparing them side-by-side, HDV showed the most macro-blocks and mosquitoes surrounding the moving objects, but these really wasn''t obvious at full frame playback. DVCPRO HD was the most block-resistant of the three formats, showing the least artifacts.

Figure 4. The Red One''s depth of field makes compression a snap.
Click here for a larger image

Along the way, I also noted that the depth-of-field performance of the Red One should make it a dream for shooting for streaming. For example, Figure 4 shows the milk girls clip that I downloaded from the Red website. Even at 5Mbps, about 20 percent the data rate of AVCHD, the clip was virtually artifact free because so little of the image was in focus. Now that I know that the workflow isn''t horrendous, it might be time to start begging for a Red camera for testing for this purpose.

Anyway, back to our format comparison. By this point, I had reached the conclusion that DVCPRO HD offered slightly higher quality than AVCHD and HDV, though at four times the data rate, it obviously comes with its own set of storage related issues. Still, this conclusion helped shift the focus to a two-horse race between full resolution AVCHD and HDV.

Figure 5. At the same data rate, HDV shows macro-block artifacts while AVCHD is smooth and clear.
Click here for a larger image

With DVCPRO HD out of the picture, I could start encoding my files progressively and start dialing down the data rate to make the MPEG-2 that fuels HDV and the H.264 behind AVCHD start feeling some pain. My rational was that I could keep downloading and encoding files from the Internet all day and still not see a significant difference at each format's full encoding parameters. However, by reducing the data rate, I could simulate the performance of the respective formats when shooting particularly challenging footage, as well as clear up an internal debate as to the superiority of H.264 over MPEG-2.

Accordingly, I produced two videos at 5Mbps using MPEG-2 at 1440x1080x24fps and H.264 at 1920x1080x24fps: Michael Hastings'' great shot of the space shuttle launch (see Figure 5) and the second the Corben clip shown in Figures 3 and 6. In the first clip, HDV shows significant blockiness as the camera tilts upward following the space shuttle, while AVCHD remains clear.

Figure 6. Ditto in this shot. No question which format you''d prefer.
Click here for a larger image

In the second clip, HDV again shows severe blockiness. AVCHD looks a bit smudged, while remaining clearly preferable to its MPEG-2 based competitor. This is particularly interesting because although I produced the AVCHD clip at full resolution (1920x1080) and the HDV clip at 1440x1080, the per-pixel bit rate of the HDV clip was 25 percent higher.

The bottom line is that a full-resolution AVCHD camera should produce better video than an HDV camera because the underlying compression format is superior. Every camera is different, of course, and optics play an obviously major role, but if all other factors are equal, AVCHD should do better.

Figure 7. I don''t often use all eight cores this efficiently, but it''s nice to know they''re there.
Click here for a larger image

One final shout-out. As late as I was on this project, I would still be working right now if not for the stunning performance of the eight-core HP xw8400 workstation I use as my primary workstation. At one point, I had RedCine converting 4K footage to QuickTime and Carbon Coder batch-encoding multiple HD clips, and I expected to see smoke coming from the machine. But like the proverbial Energizer bunny, it kept going and going. If you''re working with HD footage—affordable or not—you''re going to need an eight-core system.