Re: [codec] Comments on draft-ietf-codec-ambisonics-01

[Jan Skoglund <[e-mail]>]

Thanks Mark, Jean-Marc, and Tim,

It makes sense to use the table already defined in RFC 7845, which then
automatically allows mixed order also for family 2. We'll remove the
"channel numbering table" for family 3.

Cheers,
Jan



On Mon, Mar 13, 2017 at 4:06 PM Timothy B. Terriberry <[e-mail]>
wrote:

> Jean-Marc Valin wrote:
> > For family 3, if you already have a Cx(N+M) weight matrix, then the
> > mapping table becomes completely redundant.
>
> I think the issue boils down to whether it's worth it to have a table of
> size C to allow a matrix of size Kx(N+M) instead of Cx(N+M) in the
> headers where D is some number smaller than C (probably D=N+M). If more
> than 80% of your matrix is filled with zeros (100 out of 121 rows for
> 10th order with only horizontal channels), that could be seen as a
> little wasteful.
>
> That said, we very intentionally chose the direction of the channel
> mapping table for RFC 7845: one value per output channel that says which
> decoded channel to use. That makes it impossible to send multiple
> decoded channels (mixed channels after multiplying by the weight matrix
> in this draft) to the same speaker (ACN number in this draft). As Mark
> Harris points out, this is a failure case the current draft doesn't
> specify how to deal with. By doing the mapping in the same direction as
> RFC 7845, the worst failure case is that someone specifies a channel
> index larger than what's actually available, but that's easily
> detectable (i.e., you can tell without having to look at any other
> values in the mapping table). You also don't need to handle the diegetic
> channels separately. There is a special case for silent channels, but as
> a benefit it become easy (O(1)) to tell if a channel isn't used.
>
> Having the mapping work in this direction would also be more consistent
> with Channel Mapping Family 2.
>
>