MultiHypothesis MultiHypothesis Pictures For H.2 Pictures For H.2 6L 6L Markus Flierl Telecommunications Labor atory University of Erlangen-N uremberg Erlangen, Germany [email protected]Thomas Wiegand Image Processing Depa rtment Heinrich Hertz Instit ute Berlin, Germany [email protected]Bernd Girod Information Systems L aboratory Stanford University Stanford, CA [email protected]To be published in Proc. ICIP, Thessaloniki, Greece, Oct To be published in Proc. ICIP, Thessaloniki, Greece, Oct
22
Embed
MultiHypothesis Pictures For H.26L Markus Flierl Telecommunications Laboratory University of Erlangen-Nuremberg Erlangen, Germany [email protected].
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
MultiHypothesis PicMultiHypothesis Pictures For H.26Ltures For H.26L
To be published in Proc. ICIP, Thessaloniki, Greece, Oct. 2001To be published in Proc. ICIP, Thessaloniki, Greece, Oct. 2001
OutlineOutline• Motion Estimation in H.26L• Introdution to MultiHypothesis Pictures• Coding of MultiHypothesis Pictures• Experimental Results
Motion Estimation in H.26LMotion Estimation in H.26L• Low Complexity
– MC_Range for 16x16 Macro Blocks– Range = ½ MC_Range for Other Sub Blocks– Range = ½ Range for Search in the older ref.
pictures– MVs must be inside of the ref pictures– Choose the results which SA(T)D is minimal.
Motion Estimation in H.26LMotion Estimation in H.26L• High Complexity
– One MC_Range for all Inter Mode and Ref. Pictures– Sub Block Matching starts at the result of 16x16 Ma
cro Block Matching results– Not Forced to contain the MV (0,0)– MVs could be out of the boundaries of the Ref. Pictu
re– Choose the results which much matching the RD co
nstrain.
MultiHypothesis PicturesMultiHypothesis Pictures
MultiHypothesis PicturesMultiHypothesis Pictures• An extension of P pictures.• Each macroblock can be compensated by a linear co
mbi-nation of two motion-compensated macroblocks.
Modified in TML 6Modified in TML 6• Add a individual ref. parameter for
each block.• It’s efficient for H.263, but not for
H.26L– In H.263, 16x16 and 8x8
MacroHypothesis TypesMacroHypothesis Types
MacroHypotheses CodingMacroHypotheses Coding• In the multihypothesis mode
– Two macrohypotheses– For Each macrohypothesis
• Picture ref. Parameter• A Macrohhypothesis type• The corresponding motion vector for each sub-b
lock
EncoderEncoder• A Lagrangian cost function is used for coding
mode decisions.– D: Multihypothesis prediction error– R: Bit rate for two picture ref. Parameterstwo picture ref. Parameters, two mactwo mac
rohypotheses typesrohypotheses types, and the associated motion vecthe associated motion vectorstors.
• For multihypothesis mode– Rate-Constrained multihypothesis motion estimati
on.– Performed by the macrohypothesis selection algo.
Iterative algo. for conditional rate-coIterative algo. for conditional rate-contrained motion estimationntrained motion estimation
• Initial macrohypothesis– The best macroblock type for long-term memory m
otion-compensated prediction.• Steps
1. One macrohypothesis is fixed and conditional rate-constrained motion estimation is applied to the complementary macrohypothesis such that the multihypothesis costs are minimized.
2. The complementary macrohypothesis is fixed and the first macrohypothesis is optimized.
3. To repeat step 1, 2 until convergence.
Rate-Constrained Multi-Hypothesis Rate-Constrained Multi-Hypothesis Motion-Compensated Prediction for Motion-Compensated Prediction for Video CodingVideo Coding
M. Flierl, T. Wiegand, and B. Girod, “Rate-ConstrainedMulti-Hypothesis Motion-Compensated Prediction for VideoCoding,” in Proceedings of the IEEE International Conferenceon Image Processing, Vancouver, Canada, Sept. 2000,vol. III, pp. 150–153.
Experimental ResultsExperimental Results• Our coder is based on the H.26L test mo
del TML-6. For our experiments, the CIF sequences Mobile & CalendarMobile & Calendar and TempeTempetete are coded at 30 fps30 fps. We investigate the rate-distortion performance of multihypothesis pictures with respect to H.26L P pictures for various long-term memory buffer sizes.