Flow Motion v2.02 Low Light Comparison

Download link to latest version:

http://www.personal-view.com./talks/discussion/3337/gh2-flow-motion-v2.02-spans-in-hbr-25p-24h-720p/p1

Flow Motion v2 maintains its finest quantization quality even in dim lighting situations. These frame grabs taken from a smooth panning shot compare FM2's performance in 1080p25 HBR, 1080p 24H, and 1080p30 MJPEG HD modes. These videos were shot two-stops underexposed, illuminated by a pair of 75W incandescent bulbs. The video gain of each frame was boosted by 50% in Adobe After Effects CS5.5, and the frame grabs were cropped at 200% resolution.

Flow Motion v2 1080p 24H mode, 200% crop:

Flow Motion v2 1080p25 HBR mode, 200% crop:

Flow Motion v2 1080p30 MJPEG HD mode, 200% crop:

EDIT: Thanks to @crunchy, I discovered a footage interpretation error that caused spurious interlace artifacts to appear in the original HBR screen grab I posted above. I've replaced that with a corrected version, which displays HBR image quality virtually indistinguishable from the 24H frame grab.

It was. I added the cinegrain package to it. I absolutely love it.

@crunchy I pasted the Vimeo download link for the HBR version in between the two embedded videos in the post above. Just copy it into your browser's address bar to link to it.

Hmm, I was able to download the video from that Vimeo HBR link, even after I logged out of Vimeo. If you are unable to access it directly, you can go to my Vimeo index page and click on the HBR video:

https://vimeo.com/lpowell/videos

@LPowell, I have imported your HBR clip by using Avisynth script:

DirectShowSource("LPowell_HBR_101354524.MTS")

ConvertToRGB32

and opened it with VirtualDub. The picture is now much better (see below pure 200% resize). I've used frame 50.

I tested Flow Motion 2.01 first with the new Panny 25mm/f1.4 The only failures were with 720pSH from either Manual Movie or from Still setting pushing the red button. The subject was a large avocado tree in the wind panning slowly across it from under. One time the camera froze and I had to take the battery out to reset and no file was written in that case; other times failures gave a message and file was created up to failure spot. 720pH worked fine and pretty high quality. Using a manual lens, like Minolta 35mm/f1.8, works fine and I could not make it fail. Thus, it appears that an AF lens plus 720pSH with a subject that has lots of changing details may be too much for the camera and Flow Motion 2.01.

@cubrun I'm not an expert, but from everything I gathered here and there you can't just change the bitrate and expect everything to work fine. The AVCHD modes and parameters have some tight connections that will make it break if you operate large changes on single parameters without adjust almost everything else accordingly. What you are proposing is like driving a Ferrari at 60mph, it will run but doesn't make any sense to do it.

If you want extreme efficiency, you need to go with a long GOP setting like Sanity as it's performance/cost ratio is pretty much unrivalled. I find it's pretty good overall but some people will run away from long GOP settings like the plague.

@LPowell I can't upload now Lee (Storms again) but maybe you can shed some light on this. In both 24p and 720p I get great, almost noise free video at night.

The other day when videoing under a shelter (no sides) I got a disturbing amount using a 2.8 lens and ISO 200. 720-60, bright light outside and shaded in, shutter 1/125, I don't remember the fstop, smooth -2. This was showing up mostly in the Green.

Any ideas?

Aside from that I swear this patch is magic in 720.

@svart Flow Motion v2 is not bitrate-limited in low-light conditions - it will use as much bitrate as it needs to encode the video at its finest quantization quality. If you use Stream Parser to examine the DC QP level used in each frame, you'll see that FM2 maintains a uniform DC quantization of 5 at even the lowest levels of illumination. (Levels of quantization below 5 rapidly reach diminishing returns without perceptibly improving image quality.)

I'm not sure what you're referring to as "macroblocking", as the AVCHD encoder's inherently coarse sampling of dark shades can produce blotchy shadow effects regardless of how finely the macroblocks are quantized. Flow Motion v2 is designed to suppress high-detail chroma noise while enhancing luminance details. Since shadow areas usually have very low-contrast luminance details, they will turn out more contoured rather than noisy in FM2.

I lifted the shadow details in the Flow Motion v2 24H frame grab above, and graded it to reduce the saturation in the shadows while maintaining balanced color tones in the mids and highlights. (Yellow flowers are excellent test subjects for this purpose, as the gradients on their pedals are very sensitive to red/green color balance.) Bear in mind that the highlights in the video were intentionally recorded two-stops underexposed, as a test of low-light performance. This type of shadow grading requires very fine adjustments of individual RGB gain, gamma, and pedestal controls.

@svart The quantization tables used in Sedna are designed for an Intra-only encoder. They employ overkill techniques to record as much high-detail image data as possible, regardless of how perceptible the finest details may be in practice. As a result, Sedna records a large amount of sensor noise, as well as genuine image details, and to do so at high quality requires very high bitrates. At conservative ISO and medium and high illumination levels, sensor noise is virtually imperceptible, but in areas of very low exposure, sensor noise becomes significant. As with dithering, a layer of noise can conceal underlying digital artifacts in a way that appears to improve shadow depth.

To my eyes, sensor noise looks much better in still frames than it does in low-fps video. I find random patterns of swirling noise to be distracting rather than enhancing. With Flow Motion v2, I've specifically prioritized continuous motion picture quality over static still-frame image quality. I want to see frames that integrate smoothly over time, and for that purpose, I don't want the encoder to preserve myriads of randomly varying statistical details.

This, by the way, is why I favor Flow Motion's 3-frame GOP over Intra patches. As animators have found, the most fluid motion picture effect is achieved when one or two tweened frames are inserted between each keyframe in a video sequence. In FM2's case, the pair of B-frames that follow each I-frame are actually reconstructed using a technique similar to animation tweening. Inter-frame macroblocks in a B-frame are made of an averaged combination of motion vectors taken from both preceding and following I-frames. These tweened motion vectors are then refined with a layer of residual details extracted from the original source frame. This methodology insures a high degree of inherent frame-to-frame continuity.

With an Intra patch, each frame is recorded as a separate individual keyframe, technically uncorrelated to adjacent frames. In this case, motion picture continuity is produced by recording each frame with enough static image quality to independently preserve all details common to each sequential frame. With extremely fine levels of quantization, random discrepancies between frames (which would enhance the detail of a still image) introduce sporadically varying details that I find distracting. In my view, optimal motion picture quality is produced by careful control of image detail rather than wholesale maximization of bitrate.

Part of a Project I'm working on. FlowMotion 720p 50, 1/100, converted to 25p slowed down 50%, 14-140 kit lens. Shot from the back of a Jetski.

Best viewing the downloaded file.

Agree that you really have to look at the flow of images for IQ--you can't really see how noise behaves in a static shot. The classic example of this is pulsing--which I find really annoying. You would not see it on a still, but once the movie starts, it is really noticeable.

@LPowell

A little off-topic, but per your question on rambo's exposures -- I'm finding that with FM (but really, all patches/setting with the GH2) it's always better to expose to the right, even into slight clipping, and then bring mids and lows down in post. The files are slightly larger than when "correctly" exposed, and (more importantly) more correctable.

Is there any consensus on this question -- at least for uncontrolled lighting conditions?