I think it's a bug. Increasing overall bitrate didn't help. Lowering AQ helped. The issue goes away at AQ0. I've seen the patterns before. It happened in 66Mbps GOP6 and 88Mbps GOP6, too. 88Mbps GOP3 AQ3 didn't have any issue, but it failed high detail test.

44Mbps GOP12 AQ3 is looking good. I frame size reaching 7Mbps. B & P frames looking "healthy". Good candidate for baseline 24L setting.

So here is what I have tested so far.

I tested the firmware without any changes against 42 mb/sec High and 24 mb/sec low settings for all of the Qauntizer settings.

The only thing I varied between tests was the Quantizer. After I tested the default firmware I set the bit rates to 42 and 24 and started with a Quant of zero. Then I tested each Quant after that without changing any other setting and without adjusting the camera at all. I left all of the GOPs at their default firmware settings.

I filmed the resolution chart with my 55mm F1.2 stopped way down to make sure it was sharp enough. It was probably somewhere around F2.8. The focus, ISO, WB, Shutter Speed, and lighting was fixed. I used an external video light to keep a constant light on the resolution chart. Since I had the light and faster lens I was able to hold the ISO to 160 the whole time even indoors.

I wanted to know if increasing the Quant value would increase the bit rate. With the default firmware the bit rate was only 13.8 mb/sec for the 1080p resolution and 10.4 mb/sec for 720p. Interestingly the High and Low settings had the same bit rate for the default firmware across all resolutions with this relatively easy scene to process.

As I increased the Quant value the bit rate also increased for each of the resolutions. However, the default firmware and the firmware with Quant = 0 had the same video bit rate even though the Quant = 0 setting had the bit rate value increased to 42 mb/sec. This is interesting because raising the bit rate value alone did not help at all with this relatively simple scene.

Once I raised Quant to 1 the bit rate began to rise for all of the formats. For the 1080p High file it went from a low of 13.6(Quant = 0) through the following values 18.2(Quant = 1), 26.6(Quant = 2), 27.8(Quant = 3), 36.0(Quant = 4). All of the rest of the formats followed a similar pattern. The 1080p resolution always had the highest bit rate even though the 1080i resolution used the exact same settings with B-frames.

However, I did notice one interesting thing. Once I got to the Quant value of 2(More to details) the 720p P frames took on an uneven cadence. They began to have one large P frame followed by one small P frame. This pattern continued all the way up to Quant = 4. None of the other resolutions exhibited this pattern but the other resolutions all use B-frames.

I looked at all of the video files and I never could tell a difference in visual resolution at all. They all looked the same. I never expected it to affect resolution because even 12 mb/sec was probably plenty for this easy to process scene.

In the end I settled on Quant = 4 for 1080p because I wanted to keep a very even bit rate across all scenes no matter what level of detail was in the scene. I choose Quant = 3 for 720p even though I had seen the uneven cadence for values of 2 or greater. I did this because I lowered the GOP to 12 for 720p and 1080i. After further testing I determined that this lowering of the GOP eliminated the uneven cadence. However, I have not determined if the uneven cadence is actually a bad thing. It may be that it was uneven but still higher.

So in the end with a few simple changes I was able to get what I think is completely stable(even with a slow class 6 Eye-Fi card) video with extremely high detail and a near constant bit rate with an even cadence.

I tested the 720p GOP = 12 by shaking the camera violently with a very detailed scene. It definitely broke the codec. Several scenes showed portions of the scene in perfect focus with no Mud while the lower parts of the frame showed sever Mud and blurring with Macro Blocks. However, I never shoot like this in real life so my settings did not attempt to eliminate this.

Here is a sample 1080p Quant 4 video. This is straight out of the camera and you can download the original on Vimeo.



Here is a nice little moon video I just shot as well.



Please note that the settings file below also includes changes to the MJPEG settings. I raised the bit rate for 720p and 480p and I increased the resolution of 480p.

@cosimo_bullo My new favorite is 12 GOP on both 24p and 60p.

@mpgxsvcd the uneven cadence happens at GOP3 and GOP6, too. It goes away at GOP12.

@LukeV Based on my testing and @mpgxsvcd's testing, 44Mbps, GOP12, AQ4 for 24p, AQ3 for 60p seems to be good candidate for the baseline setting. I'm waiting for @cbrandin's feedback.

@mpgxsvcd

What do you mean by "uneven cadence" with the 720p settings? Are frames being displayed at an uneven interval? Nice report, by the way...

Chris

@cbrandin

The uneven cadence is simply the high value P frame followed by the low value P frame. You can see it in the Stream parser file for the first attachment. The standard pattern that I have seen for P frames is a very consistent amplitude that gradually raises and lowers.

I see no ill affects in the visual quality of the videos with this uneven cadence. It just seems like the bits alocated to the P frames goes up and down faster than normal. It may or may not be a good/bad thing. I am not sure. I just wanted to point out that it does happen.

I had problems with "uneven cadence" as well. It's usually like this:

Camera starts fine, with Big I-frames, followed by slightly smaller P-frames and smallest B-frames in between. Suddenly, the I-frames will get "flattened" and both I,P and B-frames get the same amplitude, or there will be saw-tooth pattern. While other didn't notice IQ-differences, I noticed strong pulsing in darker areas of detailed shots.

After lengthy testing (probably about 20 re-flashes), I "solved" the problem of uneven cadence by increasing Frame Buffer, Muxer buffer and Video buffer.

My findings so far:

Video buffer seems to contain rendered file before being written to card. Thus it will not influence bitrate. It will only safeguard uneven transfer rate to card.

"Uneven cadence" seems to happen when I-frames get so big that they cannot fit into buffer. The camera probably throws away some info to make it fit and resulting P-frames are smaller too as they are predicted from "maimed" I-frame. So you get macro-blocking, and funny looking GOP.

After raising the FL, uneven cadence went away. I assume that uneven cadence happens when I-frames cannot fit into FL-value. Here are my settings:

Video Bitrate 24H=52000000
Video Bitrate 24L=52000000
Video Bitrate FSH/SH=48000000
Video Bitrate FH/H=42000000
Auto Quantizer for 1080 modes=3 - Most to details
Auto Quantizer for 720 modes=1 - Some to details
Video buffer=0x3000000
Video buffer 24p=0x3000000
Muxer buffer for other modes=25165824
1080p24 Frame Limit=15728640
50fps Frame Limit=7549746


P.S. When I enable B-frames for 720p, the "uneven cadence" problem will not manifest itself as easily. Unfortunately, that means "1080p-dance" before each shot. Otherwise, I found that B-framed 720p is superior to IP, even at lower bitrates.

@cbrandin Why 24p AQ4 and 60p AQ3 as baseline? Good question. I don't know. I guess both can be AQ3 to keep it simple.

Anyways... it seems this effort is in freeze until the bug gets fixed.

You gotta go past 240th frame or the first 10 seconds to see the anomaly.

I don't think that "irregular cadence" has to do with EV-settings or such. -2 EV will often force camera to use lower ISO, resulting in less grain in the picture. Picture will be easier to compress so I-frames won't hit the limit, and you get working GOP.


The worst-case scenario I use is just jacking up the ISO really high, resulting in lot's of grain. It will force codec to hit it's bitrate limit trying to encode the noise. If there are "irregular cadence" problems, they will manifest themselves then.

@Grunf I used same settings. The only difference was strength of ext light. I can't conclude anything out of it. But -2 EV didn't have the irregular cadence issue. Just more clue for Chris. That's all.

@cbrandin and @rigs

The B-frames look big as well. Nearly 2 mb for a B-frame is a lot. There really isn't much compression going on in that file.

Hi I'm looking for the settings that were used on this video but I'm unclear about if they are in this thread. It looks like what @rigs posted, but without assuming anything can you guys help direct me, this video looks incredible.

Vitaliy has acknowledged the low GOP problem and is working on it. I think that we might want to put off additional low GOP testing for now.

I think I have come up with a simple way to test for better low level quality. Shoot a res chart (or some other fairly busy target) and start underexposing it until you can't see it any more. Is there a difference between settings as to how far you can go before the image is no longer visible? How does the image look when it's almost invisible - is it smooth or blotchy? These tests would be best done at ISO 160, as the camera loses dynamic range as ISO goes up and that may be the limiting factor rather than the codec.

I have found that setting levels (I use Photoshop for this on framegrabs) to bring a dark image up to normal makes things very visible (it's ugly, but you can see what's going on).

Chris