NETVC IETF 96

Administriva: Intro

Requirements

Mo: Question about optional requirements. Multiple profiles already listed, 8bit, 10bit and 12bit. Is 16-bit now another profile? Do we have a list or profiles we are going to support "right out the door" and is that list somewhere?

Alexey: There are two color formats specified in the basic requirements section; all encoders must support

Mo: So a codec candidate must support all these profiles but a decoder candidate does not need to support all of the formats?

Mo: We should also differentiate between input and output bit depth and internal representation as there are some implementations and applications where this is important.

Alexey: Yes, I agree.

Tim: Regarding if 10bit should be in a separate profile, it may be driven by what hardware can support out the gate.

<unidentified>: Do the low latency requirements apply to screencasting?

Alexey: Of course, the low delay requirement should be applicable to all applications.

Miguel Paris: A nit here about "mathematically lossless", in applications where this matters the terminology is "numerically lossless"

Alexey: I have defined these, but yes it is a good point.

Chairs: We have had this as a document in the WG for a year now, we would like to put this through WG last call in the next two weeks, so expect that.

Test and Evaluation Criteria

Chair: Are you looking for help on the rate control testing?

Thomas: Yes, it would be helpful. I need verifier software for this.

Mo: VMAF metric was added, does that support catching temporal artifacts? We see some clips with temporal jitter and the metrics are not catching that.

Thomas: Yes, VMAF does include temporal results. I am not sure how well it works, or if it will catch the issues you describe. We'd need to test that.

Andre Norton: This metric takes into account temporal effects in a local region

Mo: would it catch an artifact like pulsing?

Andre: If you see pulsing on the static scenes then it would rate this clip lower than with a fast moving scene. But it would not necessarily find htis.

Mo: Sounds like we need another metric.

Thomas: We could creata meta-metric by looking at how a metric performs over time.

Chair, to Thomas: Take that as another action item.

Thor Update

Tim: when you use a temporal motion vector, does that impact the coding if you have loss elsewhere in the frame?

Steinar: I am not sure.

Tim: This is an important question to answer as it matters for error resiliance.

Mo: We currently don't have an error resiliance flag in Thor, but we can add one.

Mo: How does this compare to the Chroma-from-Luma proposal from Mozilla?

Tim: Our original CfL implementation did a linear regression in the spatial domain, but we found that it was less reliable than doing it in the frequency domain. We actual just copy the luma coefficients and use them directly with PVQ.

Steinar: <question about intra v inter>

Tim: We currently do not use CfL on inter frames, mostly because like you we were not expecting to see any gains there, however we might have to look at that again.

Nathan: It sounds like the signal free implementation requires the decoder to also compute the linear regression for all blocks (to check against 0.5 correlation coefficient threshold). What complexity impact does this have on the decoder?

Steinar: The comparison is done without computing the square root or division, so that comparison is cheap to do.

Mo: Does this work with other planes, e.g., alpha or depth? Does this model try to predict from more than one plane?

Daala Update

Mo: Do you adapt, starting fresh each frame or do you have interframe dependencies.

Tim: Currently we reinitialize all the probabilities to a flat model.

Mo: And you don't have any flag that lets you re-use the probabilities from the previous frame?

Tim: No, in general Daala does not have any features you need to switch off for interactive video use cases.

Steinar: Does this rate control assue that there is no latency concerns? The filter would either add latency or assume that overshoot is okay for a while. But sometimes you can't overshoot.

Tim: So that is basically controlled by the buffer interval.

Steinar: How small a buffer can you go to.

Tim: The code allows for 1 frame, but it doesn't work well with 1 frame. The smallest is probably 12 frames before you need to do sub frame adjustments.

Mo: <question about adapting q>

Tim: There are tradeoffs that are better handled using temporal rate control. If you know a key frame is a good predictor of later frames, you can use more bits now. Likewise if your keyframe has a bunch of motion that only needed now, you could code it with a similar quantizer as a P-frame.

Mo: Could estimates of available bitrate be modelled and effect the rate control.

Tim: Yes, the algorithm does a budget allocation when encoding each frame and then throws it away after encoding a frame. This means you can adapt quickly to changes in network environment.

Administriva: Wrapup

Chair: Good progress on the two candidates. It may be useful now to see how they compare to our anchors [codecs].

Ramalho: Is there a schedule or timeline for testing? There is a date in 2017.

Chair: Its a poorly worded milestone. The milestone is for the delivery of the testing draft, not test results.