Flow Motion v2 - Minimization of Macroblock Artifacts

A flowing stream of water can require unexpectedly high bitrates to encode with visually flawless image quality. To an AVCHD encoder, each wave contains a myriad of independent motion vectors that fluctuate in unpredictable patterns. An encoder that handles highly randomized motion gracefully, without producing noticeable digitizing artifacts, is well equipped for a wide range of subjects. Here is an unedited comparison of the unhacked GH2's AVCHD encoder in its highest quality 1080p 24H mode with Flow Motion v2:

Unhacked GH2 in 24H Mode

Flow Motion v2 in 24H Mode

An unhacked GH2 actually looks pretty good on consumer-level equipment. For post-production grading or theater projection, however, magnification reveals unsightly macroblock artifacts. In the 200% crop frame grabs below, the upper image is from the unhacked GH2, while the lower image was produced by Flow Motion v2:

(I've also included unedited frame grabs at 100%, 200%, and 400% magnification below.)

One of the major causes of AVCHD macroblock artifacts is the common industry practice of encoding P and B-frames with coarser quantization tables than are used in I-frames. This enables professional Blu-ray encoders to pack lengthy movies into optimally compact files for commercial distribution to end-users. For capturing original source footage, however, this encoding strategy can compromise motion picture quality with painfully visible artifacts of distinctly digital characteristics that cannot easily be corrected in post production.

Macroblock artifacts not only look ugly, they can actually sabotage the encoder's best efforts to accurately encode P and B-frames. This is because the reference frames used as the basis for encoding P and B-frames are not the original images, but are instead decoded images reconstructed by the encoder's built-in Deblocking Filter. When those reference frames are contaminated with coarse macroblock artifacts, the encoder must then squander much of its available bitrate to suppress the propagated artifacts from reappearing in subsequent P and B-frames.

Unlike DSLR's with hardware-based AVCHD encoders, the GH2's software encoder works under an additional constraint. While hardware encoders can use both 4x4 and 8x8 quantization matrices to encode individual macroblocks, the GH2 is limited to using only 4x4 matrices. This puts it at a disadvantage when encoding subtle low-detail gradients, which 8x8 matrices can render with higher quality at low bitrates.

To compensate for these technical shortcomings, the GH2 needs significantly finer P and B-frame quantization tables that minimize macroblock artifacts and match the image quality of its I-frames. In addition, it requires significantly higher bitrates to utilize these tables at the finest practical levels of quantization. This is what the Quantization and Deblocking Tables developed for Flow Motion v2 are designed to accomplish.

GH2 AVCHD Encoder Failure Patterns

Most DSLR AVCHD encoders have been implemented at least partially with dedicated encoding hardware to insure consistent image quality and performance in all video operating modes. This includes the Panasonic AF100, which uses a DSP implementation that delivers perceptibly better quality than the unhacked GH2 within the constraints of AVCHD's 24Mbps maximum bitrate specifications.

As with all Panasonic G-series cameras, the GH2's AVCHD encoder is implemented entirely in software. While this helps reduce the manufacturing cost of the camera, it makes the encoder's real-time performance vulnerable to disruption by other demands that may be placed on the camera's CPU while shooting video. In particular, automated mechanisms such as OIS, dynamic auto-exposure, and continuous auto-focus require constant attention to changing conditions which may transiently undermine the encoder's real-time performance.

With a fixed frame rate to maintain, the AVCHD encoder has but a limited amount of time in which to encode each batch of frames and write them out to the SD card. It must also work within the limits of the peak bitrate available to encode each frame. If at any time the data size or encoding time requirements of a frame exceed the encoder's limitations, it may resort to operating in a built-in low-quality Fallback Mode, which uses coarse quantization to quickly encode the frame before its time runs out.

Although the GH2's software implementation allows the AVCHD's bitrate and image quality to be greatly improved, this also increases the hazards of the camera failing under demanding conditions. Unlike SD card write-speed errors that cause video recording to abruptly halt, Fallback Mode failures occur silently with no apparent disruption of the recording process. The degradation in image quality caused by Fallback Mode may not always be immediately noticeable, raising the possibility that these failures may insidiously creep into footage that otherwise appears to be recorded at the intended level of quality.

Unfortuntely, Fallback Mode can occur even when shooting ordinary scenery with the GH2's unhacked 24Mbps AVCHD firmware. Here is an example shot in 1080p 24H mode with a Panasonic 45-200mm lens:

At first glance this video looks fine, and it is difficult to spot which frames have been recorded with degraded image quality. In these cases, cbrandin's invaluable Stream Parser may be used to examine the encoding quality of each frame in detail. The Elementary Stream Decoder built into Stream Parser directly identifies frames encoded in Fallback Mode, and reports a numerical rating of each frame's quantization quality. Here is an extract of the Stream Parser report on the video above:

In the bar chart at the top of the report, each GOP begins with a red I-frame. In GOP's 2-4, a bitrate disruption can be seen which provokes the encoder to plunge into Fallback Mode in the following GOP's. The chart at the bottom lists the quantization factors used in each recorded frame, along with a quality rating under the "DC" column. In the "QST-High" column, each frame encoded in Fallback Mode is marked with an asterisk.

The DC values of 26-39 used in Fallback Mode indicate a very coarse level of quantization, degrading the image quality of those frames. (Larger values indicate lower quality.) After the encoder recovers from Fallback Mode, it maintains significantly higher quality DC factors of 9-13, a reasonable level of quality to expect from the AVCHD encoder's 24Mbps maximum bitrate.

Using custom rate control techniques, Flow Motion v2 is able to sustain peak bitrates up to 100Mbps with complete protection from Fallback Mode. It is designed to maintain consistently superior quantization quality over a broad range of shooting conditions, including extreme underexposure by up to two stops. In real-life shooting conditions, Flow Motion v2 encodes at DC values commonly ranging from 5-9, the finest levels of quantization the AVCHD encoder can produce before hitting rapidly diminishing improvements in perceptible image quality at DC values of 4 and below. The combination of excellent motion picture quality with efficient 100Mbps peak bitrate consumption in all video modes makes Flow Motion v2 one of the highest performance AVCHD patches developed for the GH2.

@ LPowellLPowell 11:18AM.... thank you very much Lpowell. I'll start to test immediately...

@LPowell

To use, unzip the INI file into the same folder as the PTool application. Launch PTool and load the firmware v1.1 binary file for GH2. The settings contained in the Flow Motion v2.0 INI file will automatically be installed in the "F" button at the bottom of the PTool main window.

Want to remind that for long time PTool can install directly zip or plain ini patches from any folder.

just done, Rokkor 50mm 1,4 - shutter 50, f2,8 iso 160, inside, on tripod, screenshot results,

@LPowell

Just downloaded and will be loading this up on my second GH2. I really liked your FM 1.1 but have since been using Sanity 5 for it's incredibly small files sizes and reliability.

I'll try to do some side by side testing just for the heck of it, but I'm sure I'll probably just end up keeping one camera loaded with each.

Thanks for all your hard work, all you developers (too many to name...) are a great bunch of guys for not only taking the time to sift through and experiment with all this data, but also for sharing it so freely with all the rest of us consumers. Hats off!

Tony

i also use sanity 5. Just make a small test, and I THINK that with your patch the video is brighter!!!? It catches more light on same conditions!!! ??? As i understand it is not span in HBR mode, right ?

@matthere 1:44PM ...why is your result so much different to my 24p (1. and 3rd screen)

First impression: W-O-W !!!

@matthere 2:09PM ... I mean the bad streamparser result of my shots, comparing your clear/straight/perfect results..

The subject will have an effect, not sure what you had in your sights, but bit rate will vary, the more complex the scene you are shooting the greater the bitrate will be.. as for the clarity or straightness of the lines.. not sure on that one..

@LPowell Will test this as B cam on a commercial shoot today. I have a quick question: what exactly do you mean when you wrote this line above:

o Access to both PAL 25/50p and NTSC 30/60p frame rates.

I assume you're referring to 720p50 and 720p60 respectively, right?

@LPowell, Thanks so much for all the hard work on researching the encoder and for the informative description. Looking forward to testing this out.