Opened 8 years ago

Closed 6 years ago

#92 closed defect (fixed)

The different PartNum of different layers has problem

Reported by: kenhuang Owned by: Vadim
Priority: minor Milestone: SHM-11.0
Component: SHM software Version: SHM-10.0
Keywords: Cc: Vadim, jct-vc@…

Description

Considering the cases that the PartNum of layer0 is different to that of Layer1 1. g_auiRasterToZscan makes incorrect AbsPartIdx

  1. the mvp candidate is incorrect

Change history (6)

comment:1 Changed 8 years ago by kenhuang

Considering the cases that the PartNum of layer0 is different to that of Layer1
a.g_auiRasterToZscan makes incorrect AbsPartIdx

  1. the mvp candidate will be incorrect

comment:2 Changed 8 years ago by Vadim

I think you should enable LAYER_CTB macro for that case. Could you try it in your experiment? I fixed some compilation errors for that macro with rev 1520.

comment:3 Changed 8 years ago by kenhuang

The problem of g_auiRasterToZscan is fixed. But the problem of mvp candidate is still there.
SHM does not correctly handle the mv information of the collocated pictures. SHM scales the mv information of BL based on the PartNum of EL rather than that of BL. Therefore, the mvp candidates will be wrong when decoding EL. Considering the following case:
BL

mv00 mv00 mv02 mv02
mv00 mv00 mv02 mv02
mv08 mv08 mv10 mv11
mv08 mv08 mv14 mv15

EL

mv mv
mv x

and the Scaled BL is

mv00 mv02
mv08 mv10

Considering the marked position, x, in EL, that gets the wrong mvp candidate.

Last edited 8 years ago by kenhuang (previous) (diff)

comment:4 Changed 8 years ago by Vadim

I am not sure I understand the example, could you please point out the problematic code and possibly the fix if you have one?

comment:5 Changed 6 years ago by forayr

This issue is still present.

In TComPic::copyUpsampledMvField, the upsampled picture used the partition size of the enhanced layer. Therefore, when the partition size (widthInPu and heightInPu) is 32, we get a unitNum of 1, and motion information are derivated for each 32x32 partition.
However in the specification (H.8.1.4.3), the motion information are always derivated for each 16x16 partition.

In the previous example, the SHM will get only one 32x32 partition containing mv00.
When the upsampled picture is used to derivate the collocated center candidate (this corresponds to the position x in the example) the shm will use mv00 instead of mv10 as in the specification.

comment:6 Changed 6 years ago by Vadim

  • Resolution set to fixed
  • Status changed from new to closed

the issue was fixed with rev 1600

Note: See TracTickets for help on using tickets.

This list contains all users that will be notified about changes made to this ticket.

These roles will be notified: Reporter, Owner, Subscriber, Participant

  • forayr(Participant)
  • Gerhard Tech(Always)
  • jct-vc@…(Subscriber)
  • Karsten Suehring(Always)
  • kenhuang(Reporter, Participant)
  • Vadim Seregin(Owner, Subscriber, Participant)