Modern digital acquisition, post and distribution wouldn’t be possible without data rate reduction, AKA compression. People like to disparage compression, but I dare say that few folks – including most post production professionals – have actually seen much uncompressed content. In fact, by the time you see a television program or a digitally-projected movie it has passed through at least three, four or more different compression algorithms – i.e. codecs.
Avid Media Composer and Apple Final Cut Pro dominate the editing landscape, so the most popular high-end HD codecs are the Avid DNxHD and Apple ProRes 422 codec families. Each offers several codecs at differing levels of compression, which are often used for broadcast mastering and delivery. Apple and Avid, along with most other NLE manufacturers, also natively support other camera codecs, such as those from Sony (XDCAM-HD, HD422, EX) and Panasonic (DVCPRO HD, AVC-Intra). Even these camera codecs are being used for intermediate post. I frequently use DVCPRO HD for FCP jobs and I recently received an edited segment as a QuickTime movie encoded with the Sony EX codec. It’s not a question of whether compression is good or bad, but rather, which codec gives you the best results.
Click on the above images to see an enlarged view. (Images from Olympus camera, prior to NLE roundtrip. Resized from original.)
I decided to test some of these codecs to see the results. I started with two stills taken with my Olympus C4000Z – a 4MP point-and-shoot digital camera. These images were originally captured in-camera as 2288-pixel-wide JPEGs in the best setting and then – for this test – converted to 1920×1080 TIFFs in Photoshop. My reason for doing this instead of using captured video, was to get the best starting point. Digital video cameras often exhibit sensor noise and the footage may not have been captured under optimum lighting conditions, which can tend to skew the results. The two images I chose are of the Donnington Grove Country Club and Hotel near Newbury, England – taken on a nice, sunny day. They had good dynamic range and the size reduction in Photoshop added the advantages of oversampling – thus, very clean video images.
I tested various codecs in both Avid Media Composer 4.0.5 and Apple Final Cut Pro 7. Step one was to import the images into each NLE. In Avid, the conversion occurs during the import stage, so I set my import levels to RGB (for computer files) and imported the stills numerous times in these codecs: 1:1 MXF (uncompressed), DNxHD145, DNxHD220, DNxHD220x, XDCAM-EX 35Mbps and XDCAM-HD422 50Mbps. In Final Cut Pro, the conversion occurs when files are placed on the timeline and rendered to the codec setting of that timeline. I imported the two stills and placed and rendered them onto timelines using these codecs: Apple 8-bit (uncompressed), ProRes LT, ProRes, ProRes HQ, DVCPRO HD and XDCAM-EX 35Mbps. These files were then exported again as uncompressed TIFFs for comparison in Photoshop. For Avid, this means exporting the files with RGB levels (for computer files) and for FCP, using the QuickTime Conversion – Still Image option (set to TIFF).
Note that in Final Cut Pro you have the option of controlling the import gamma settings of stills and animation files. Depending on the selection (source, 1.8, 2.20, 2.22) you choose, your video in and back out of Final Cut may or may not be identical to the original. In this case, choosing “source” gamma matched the Avid roundtrip, whereas using a gamma setting of 2.2 resulted in a darker image exported from FCP.
Click on the above images to see an enlarged view.
You’ll notice that in addition to various compressed codecs, I also used an uncompressed setting. The reason is that even “uncompressed” is a media codec. Furthermore, to be accurate, compression comparisons need to be done against the uncompressed video image, not the original computer still or graphic. There are always going to be some changes when a computer file is brought into the video domain, so you can’t fairly judge a compressed video file against the original photo. Had I been comparing video captured through a hardware card, then obviously I would only have uncompressed video files as my cleanest reference images.
I lined up the exported TIFFs as Photoshop layers and generated comparisons by setting the layer mode to “difference”. This generates a composite image based on any pixel value that is different between the two layers. These difference images were generated by matching a compressed layer against the corresponding Avid or FCP uncompressed video layer. In other words, I’m trying to show how much data is lost when you use a given compressed codec versus the uncompressed video image. Most compression methods disproportionately affect the image in the shadow areas. When you look at a histogram displaying these difference results, you only see levels in the darkest portion of an 8-bit scale. On a 0-255 range of levels, the histogram will be flat down to about 20 or 30 and then slope up quickly to a spike at close to 0.
This tells you that the largest difference is in the darkest areas. The maximum compression artifacts are visible in this range. The higher quality codecs (least compressed), exhibit a smaller histogram range that is closer to 0. The more highly-compressed codecs have a fatter range. This fact largely explains why – when you color grade highly compressed camera images – compression artifacts become quite visible if you raise black or gamma levels.
The resulting difference images were then adjusted to show artifacts clearly in these posted images. By adjusted, I mean changing the levels range by dropping the input white point from 255 to 40 and the output black point from 0 to 20. This is mainly for illustration and I want to reiterate that the normal composite images DO NOT look as bad as my adjusted images would imply. In fact, if you looked at the uncorrected images on a computer screen without benefit of a histogram display, you might think there was nothing there. I merely stretched the available dynamic range for demonstration purposes.
Of these various codecs, the Apple DVCPRO HD codec shows some extreme difference results. That’s because it’s the only one of these codecs that uses horizontal raster scaling. Not only is the data compressed, but the image is horizontally squeezed. In this roundtrip, the image has gone from 1920-pixels-wide (TIFF) to 1280 (DVCPRO HD) back to 1920 (exported TIFF). The effects of this clearly show in the difference image.
Click on the above images to see an enlarged view.
There are a couple of other things you may notice, such as level differences between the Avid and Apple images and between each of these and the originals. As I said before, there will always be some differences in this sort of conversion. Plus, Apple and Avid do not handle color space, level and gamma mapping in the same way, so a round trip through each application will yield slightly different results. Generally, if 2.2 gamma is selected for imported stills, the Apple FCP image will have a bit more contrast and somewhat darker shadow areas when compared to Avid on a computer screen – even when proper RGB versus Rec. 709 settings are maintained for Avid. This is mainly a result of the various QuickTime and other conversions going on.
If I were to capture video with Avid DX hardware on the Media Composer and AJA, Matrox or Blackmagic hardware on FCP – and compared these images on a video monitor and with scopes – there would likely be no such visible difference. When I used “source” gamma in FCP, then the two matched each other. Likewise, when you review the difference images below, 2.2 gamma in this case resulted in a fault difference composite between the FCP uncompressed and the original photo. The “source” gamma version more closely resembles the Avid result and is the right setting for these images.
The take-away from these tests should be that the most important comparisons are those that are relative, i.e. “within species”. In other words, how does ProRes LT compare to ProRes HQ or how does DNxHD 145 compare to DNxHD 220x? Not, how an Avid export compares with a Final Cut export. A valid inter-NLE comparison, however, is whether Avid’s DNxHD220x shows more or less compression artifacts than Apple’s ProRes HQ.
I think these results are pretty obvious: Higher-data-rate codecs (less compression) like Apple ProRes HQ or Avid DNxHD 220x yield superb results. Lower-date-rate codecs (more compression) like XDCAM-EX yield results that aren’t as good. I hope that arming you with some visible evidence of these comparisons, will help you better decide what post trade-off to use in the future.
(In case you’re wondering, I do highly recommend the Donnington Grove for a relaxing vacation in the English countryside. Cheers!)
Click on these images to see an enlarged view.
©2010 Oliver Peters