Back to Search Document details
13th Meeting: Marrakech, January 2019 2019-01-17 10:17
Non-CE4: On Temporal Motion Buffer Compression
Authors: F. Bossen Sharp Labs of America Kiran Misra Sharp Labs of America Andrew Segall Sharp Labs of America
Abstract
This contribution proposes a method to reduce the number of bits stored in the temporal motion buffer. It is reported that the current VVC WD requires 22.5 bits of storage per 4x4 luma sample block, and it is further asserted that the proposed approach reduces the amount requirement to 10.5 bits per 4x4 luma sample block with negligible impact on coding efficiency. The storage reduction is achieved by first scaling the original temporal motion vector to a fixed POC distance, eliminating the need to store delta POC information. The approach then converts the scaled motion vector value to a 6-bit mantissa and 4-exponent representation prior to storage. The selected representation effectively quantizes larger motion vector values more coarsely while retaining higher precision for smaller motion vectors. The proposed method was implemented in VTM-3.0 and its performance evaluated using Random Access (RA) and Low Delay B (LDB) configurations from the Common Test Conditions. It is reported that the proposed method results in an average luma BD Rate of 0.00% (RA), 0.02% (LDB).
Meeting Notes
JVET-M0512 Non-CE4: On Temporal Motion Buffer Compression [F. Bossen, K. Misra, A. Segall (Sharp Labs of America)]

Notes from BoG report review: 0.00% (99%, 100%)/0.02% (100%, 102%). Remove POC storage and store MV in a converted way.

From subsequent discussion in Track B:

This approach stores the MV 8x8 grid in a floating point format, 6 bit mantissa (incl. sign) and 4 bit exponent. This has been cross-checked, text is available, and it provides significant reduction of MV memory. It comes with practically no loss and is well understood.

Decision: Adopt JVET-M0512 second aspect as described above.

JVET-M0512 Non-CE4: On Temporal Motion Buffer Compression [F. Bossen, K. Misra, A. Segall (Sharp Labs of America)]

Notes from BoG report review: 0.00% (99%, 100%)/0.02% (100%, 102%). Remove POC storage and store MV in a converted way.

From subsequent discussion in Track B:

This approach stores the MV 8x8 grid in a floating point format, 6 bit mantissa (incl. sign) and 4 bit exponent. This has been cross-checked, text is available, and it provides significant reduction of MV memory. It comes with practically no loss and is well understood.

Decision: Adopt JVET-M0512 second aspect as described above.

References:
JVET-L1001 JVET-L1010 JVET-M0479
Cited:
JVET-M0024 JVET-M0764 X JVET-M1001 JVET-N0002 JVET-N0003 JVET-N0075 JVET-N0076 JVET-N0077 JVET-N0078 JVET-N0079 JVET-N0081 JVET-N0084 JVET-N0085 JVET-N0086 JVET-N0088 JVET-N0097 JVET-N0107 JVET-N0108 JVET-N0115 JVET-N0116 more than 20
ARCHIVE v4 :
._JVET-M0512 - disablePOCScaling r1.xlsm ._JVET-M0512--VVC-WD-Changes_r2.docx ._JVET-M0512.pptx ._JVET-M0512.xlsm ._JVET-M0512_r2.docx JVET-M0512 - disablePOCScaling r1.xlsm JVET-M0512--VVC-WD-Changes_r2.docx JVET-M0512.pptx JVET-M0512.xlsm JVET-M0512_r2.docx
Decisions
adopted
Pending on JVET-M0170 (which was adopted)
adopted
Adopt JVET-M0512 second aspect as described above
Citation