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

[Jan Skoglund <[e-mail]>]

Ha, sorry, wrong name! I meant avoiding permutation matrices.

Jan

On Mon, Mar 13, 2017 at 3:19 PM Jean-Marc Valin <[e-mail]> wrote:

> On 13/03/17 06:17 PM, Jan Skoglund wrote:
> > Our idea was to avoid a mapping table, potentially sparse, completely
> > for family 2, and replacing it with a channel numbering list for family
> 3.
>
> Can you explain what you mean here by "avoid a mapping table" for family
> 2 and "channel numbering list" for family 3?
>
>         Jean-Marc
>
> > Cheers,
> > Jan
> >
> > On Mon, Mar 13, 2017 at 3:12 PM Jean-Marc Valin <[e-mail]
> > <mailto:[e-mail]>> wrote:
> >
> >     On 13/03/17 06:04 PM, Drew Allen wrote:
> >     > so just to be clear, if a user, say, wants to encode some mixed
> order
> >     > ambisonics using ch253, how does the decoder know what ambisonic
> >     > channels it has received and know how to render them correctly?
> >
> >     Well, each line of the matrix would correspond to a channel in the
> >     ambisonics channel order. If that channel isn't encoded, then the
> line
> >     would have only zeros.
> >
> >     The only way to avoid that situations would be to encode a separate D
> >     value (D <= C) for the number of non-zero channels among the C
> >     ambisonics channels possible. Then you'd store C values in the
> channel
> >     mapping array (equivalent to a CxD permutation matrix), followed by a
> >     Dx(M+N) weight matrix that would no longer have entire lines of
> zeros.
> >     The result would be more compact in the case of sparse
> representation,
> >     but IMO it'd be pretty ugly and prone to implementation errors. And
> if
> >     you force D==C and don't code the D (which is what I'm proposing),
> then
> >     the channel mapping permutation automatically becomes redundant.
> >
> >     Cheers,
> >
> >             Jean-Marc
> >
> >     > On Mon, Mar 13, 2017 at 3:00 PM Drew Allen <[e-mail]
> >     <mailto:[e-mail]>
> >     > <mailto:[e-mail] <mailto:[e-mail]>>> wrote:
> >     >
> >     >     Got it. In that case, it certainly seems reasonable if I
> >     understand
> >     >     correctly. Thanks for clearing that up!
> >     >
> >     >     On Mon, Mar 13, 2017 at 2:55 PM Jean-Marc Valin
> >     <[e-mail] <mailto:[e-mail]>
> >     >     <mailto:[e-mail] <mailto:[e-mail]>>>
> wrote:
> >     >
> >     >         On 13/03/17 05:44 PM, Drew Allen wrote:
> >     >         > I think the issue is that the number of total channels
> rises
> >     >         > quadratically in respect to the ambisonic order (N +
> >     1)^2. If
> >     >         a user
> >     >         > wants to use just the horizontal channels, it is only 2
> >     * N +
> >     >         1. If they
> >     >         > wish to code very high-order (>10th order) horizontal
> >     >         channels, they
> >     >         > would be artifically limited by all the zero channels
> being
> >     >         produced,
> >     >         > no? Or can this handled without actually creating all
> those
> >     >         empty channels?
> >     >
> >     >         As far as I understand, the current draft already has all
> the
> >     >         limitations you're describing. The channel mapping array is
> >     >         basically
> >     >         equivalent to a CxC permutation matrix that multiplies the
> >     Cx(N+M)
> >     >         weight matrix. The result is still a Cx(N+M) matrix, so
> >     using the
> >     >         resulting matrix as weights can still do everything
> >     without the
> >     >         need for
> >     >         the channel mapping to do the permutations.
> >     >
> >     >         Cheers,
> >     >
> >     >                 Jean-Marc
> >     >
> >     >         > On Mon, Mar 13, 2017 at 2:41 PM Mark Harris
> >     >         <[e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>
> >     >         > <mailto:[e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>>> wrote:
> >     >         >
> >     >         >     On Mon, Mar 13, 2017 at 10:38 AM, Jan Skoglund
> >     >         <[e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>
> >     >         >     <mailto:[e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>>> wrote:
> >     >         >     > Hey,
> >     >         >     >
> >     >         >     > Thanks for your comments
> >     >         >     >
> >     >         >     > On Mon, Mar 13, 2017 at 10:08 AM Mark Harris
> >     >         <[e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>
> >     >         >     <mailto:[e-mail]
> >     <mailto:[e-mail]> <mailto:[e-mail]
> >     <mailto:[e-mail]>>>>
> >     >         wrote:
> >     >         >     >>
> >     >         >     >> On Fri, Feb 17, 2017 at 1:57 PM, Jean-Marc Valin
> >     >         >     <[e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>
> >     >         <mailto:[e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>>>
> >     >         >     >> wrote:
> >     >         >     >> > 3.2.  Channel Mapping Family 3
> >     >         >     >> >
> >     >         >     >> > I would suggest removing the "Output Channel
> >     >         Numbering" field
> >     >         >     because it
> >     >         >     >> > is fully equivalent to simply permuting lines
> >     of the
> >     >         matrix.
> >     >         >     Also, I
> >     >         >     >> > believe that the size of the matrix was meant
> to be
> >     >         "32*(N+M)*C
> >     >         >     bits"
> >     >         >     >> > rather than "32*N*C bits".
> >     >         >     >>
> >     >         >     >> To expand on this a bit, a mapping family maps M+N
> >     >         decoded channels
> >     >         >     >> (corresponding to the actual order of the coupled
> and
> >     >         uncoupled
> >     >         >     >> channels in the bitstream) to C output channels
> >     >         (channels with a
> >     >         >     >> specific semantic meaning).  The additional
> >     "Output Channel
> >     >         >     Numbering"
> >     >         >     >> table confuses things by adding an additional
> mapping
> >     >         from the output
> >     >         >     >> channel numbers to a different set of numbers with
> >     >         actual semantic
> >     >         >     >> meaning, leaving the output channel numbers with
> no
> >     >         apparent meaning.
> >     >         >     >>
> >     >         >     >> This does have a potential benefit as a matrix
> >     >         compression technique,
> >     >         >     >> to reduce the size of the matrix when it would
> >     contain
> >     >         rows that are
> >     >         >     >> all zero.  However considering that the matrix
> occurs
> >     >         only once, and
> >     >         >     >> mapping family 2 already offers a way to compress
> the
> >     >         matrix, this
> >     >         >     >> alone does not seem worth the complexity of
> another
> >     >         level of
> >     >         >     >> indirection.  If matrix compression is desired it
> >     would
> >     >         probably be
> >     >         >     >> less confusing to describe it in those terms and
> keep
> >     >         the semantic
> >     >         >     >> meaning tied to the output channels.
> >     >         >     >>
> >     >         >     >>
> >     >         >     >> The description of the Output Channel Numbering
> also
> >     >         does not specify
> >     >         >     >> the intended behavior if the same value appears
> >     in the
> >     >         table multiple
> >     >         >     >> times.
> >     >         >     >>
> >     >         >     >> Additionally, section 4.2 describes how to
> >     perform a stereo
> >     >         >     downmix of
> >     >         >     >> mapping family 3, but makes assumptions about the
> >     >         output channel
> >     >         >     >> numbering.  This seems harmful and likely to
> promote
> >     >         implementations
> >     >         >     >> that make similar assumptions.  If it is
> necessary to
> >     >         apply the
> >     >         >     output
> >     >         >     >> channel numbering described in section 3.2 in
> >     order to
> >     >         implement a
> >     >         >     >> correct stereo downmix, then it would be better to
> >     >         simply use the
> >     >         >     >> output channels from section 3 as input to the
> >     downmix,
> >     >         consolidating
> >     >         >     >> sections 4.1 and 4.2, rather than specify new
> >     formulas
> >     >         that make
> >     >         >     >> assumptions about the mapping.  That would also
> >     greatly
> >     >         simplify
> >     >         >     >> section 4.
> >     >         >     >>
> >     >         >     >> Eliminating the Output Channel Numbering table as
> >     >         Jean-Marc suggests
> >     >         >     >> should resolve these concerns.
> >     >         >     >
> >     >         >     >
> >     >         >     > The problem is that once we allow mixed orders
> >     there is
> >     >         no unique
> >     >         >     way for a
> >     >         >     > receiver/decoder
> >     >         >     > to resolve the mapping to ACNs from just a number
> of
> >     >         total output
> >     >         >     channels.
> >     >         >
> >     >         >
> >     >         >     In mapping family 2, the channel count (C) is the
> number
> >     >         of channels
> >     >         >     in the fully periphonic configuration, but it is not
> >     >         necessary to
> >     >         >     encode them all.  The channel mapping table can map
> each
> >     >         ACN to a
> >     >         >     specific decoded channel or to silence.  For mixed
> >     order,
> >     >         some of the
> >     >         >     ACNs will be mapped to silence and will not be
> encoded.
> >     >         >
> >     >         >     In mapping family 3, the matrix can do everything
> >     that the
> >     >         channel
> >     >         >     mapping table can do and more.  Why not treat C in
> the
> >     >         same manner, as
> >     >         >     the number of channels in the fully periphonic
> >     >         configuration, even if
> >     >         >     some are silent?
> >     >         >
> >     >         >      - Mark
> >     >         >
> >     >         >     _______________________________________________
> >     >         >     codec mailing list
> >     >         >     [e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>
> >     >         <mailto:[e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>>
> >     >         >     https://www.ietf.org/mailman/listinfo/codec
> >     >         >
> >     >         >
> >     >         >
> >     >         > _______________________________________________
> >     >         > codec mailing list
> >     >         > [e-mail] <mailto:[e-mail]>
> >     <mailto:[e-mail] <mailto:[e-mail]>>
> >     >         > https://www.ietf.org/mailman/listinfo/codec
> >     >         >
> >     >
> >
> >
> >
> > _______________________________________________
> > codec mailing list
> > [e-mail]
> > https://www.ietf.org/mailman/listinfo/codec
> >
>