• Slide 1
  • Tomorrow: Uplink Video Transmission Today: Downlink Video Broadcast Changing Landscape of Multimedia Applications
  • Slide 2
  • Motion-Compensated Predictive Coding (MPEG/H.26) High compression efficiency Rigid complexity partition between encoder (heavy) & decoder (light) High fragility to transmission losses Image Coding (Motion JPEG) Low complexity High robustness to transmission losses Low compression efficiency + Previous frame Current frame DFD (Displaced Frame Difference) Motion search range Motion Vector Contemporary Video Coding Standards
  • Slide 3
  • Challenges: Low bandwidths high compression efficiency Limited handheld battery power low end-device complexity Lossy wireless medium robustness to transmission losses Rethinking Video Over Wireless High compression efficiency Flexible partition of complexity between encoder & decoder Inbuilt robustness to channel loss Backward compatibility with existing video standards Puri & Ramchandran, Allerton 02 Light PRISM Uplink Encoder Light PRISM Downlink Decoder Heavy PRISM Downlink Encoder Heavy PRISM Uplink Decoder Trans-coding Proxy New Architecture: PRISM (Power-efficient, Robust, hIgh-compression Syndrome-based Multimedia coding)
  • Slide 4
  • DecoderEncoder X Y X ^ X and Y are correlated sources Y is available only at decoder Source Coding with side-information (SlepianWolf, Wyner-Ziv) Exploit side-information Y at the decoder while encoding X No MSE performance loss over case when Y is available at both encoder and decoder when innovations is Gaussian For the video coding case, X is the block to be coded and the side-information Y consists of the previously decoded blocks in the frame memory Background: Distributed Source Coding
  • Slide 5
  • The encoder does not have access to Y1, Y2, etc Neither the encoder nor the decoder knows the correct side- information Can decoding work? Yes! A modified Wyner-Ziv paradigm is needed (Ishwar, Prabhakaran, & Ramchandran ICIP 03.) Predictive Decoder Predictive Encoder Quantized DFD X... Y1Y1 YMYM Y1Y1 YMYM Motion Vector X PRISM Decoder PRISM Encoder X... Y1Y1 YMYM Y1Y1 YMYM X ? Motion-Free Encoding?
  • Slide 6
  • X Wyner-Ziv Encoder bin index Y1Y1 Wyner-Ziv Decoder X YTYT Wyner-Ziv Decoder... YMYM Wyner-Ziv Decoder... Decoding failure Decoding failure PRISM Robustness Comparisons: Predictive Coding: channel errors lead to prediction mismatch and drift PRISM: drift stopped if syndrome code is strong enough: Targeted noise Correlation Noise + Induced Channel Noise + Quant. Noise Need concept of motion compensation at decoder! Need mechanism to detect decoding failure In theory: joint typicality (statistical consistency) In practice: use CRC
  • Slide 7
  • Secondary description of video sent over auxiliary-channel. Need to find statistics of correlation noise Z = X X main. Can leverage algorithm of Zhang, Regunathan and Rose (Asilomar 99) to develop recursive correlation estimation algorithm. (Wang, Majumdar, Ramchandran, and Garudadri: PCS 04.) Auxiliary channel allows drift correction without intra-refresh. Standards-Compliant Auxiliary-Channel MPEG/H.26x Encoder X Auxiliary-Channel Encoder MPEG/H.26X Decoder Auxiliary-Channel Decoder X main Final reconstruction Coset Index Wireless Channel Auxiliary-Channel X ^ MPEG/H.26x bit-stream Wireless Channel
  • Slide 8
  • Results Channel simulator provided by Qualcomm Inc. conforming to a CDMA 2000 1x standard. Performance comparison among 3 systems: H.263+ bitstream with 20% extra rate for FEC (RS codes) H.263+ bitstream with 20% extra rate for standard-compliant auxiliary channel PRISM Standard-compliant auxiliary channel version outperforms H.263+FEC by 2.5-4 dB between error rates of 2-10%. PRISM outperforms H.263+FEC by 6-8 dB between error rates of 2-10%. H.263+ with FEC Stefan, 352x240, 15fps, 2200 kbps, 8% error rate H.263+ with Auxiliary ChannelPRISM
  • Slide 9
  • PRISM for Wireless Video Broadcast Broadcast source coding studied in information theory literature. (Heegard & Berger, IT85, Steinberg & Merhav IT04) Lossy channel: need broadcast source-channel coding view. Can use PRISM constructions. (Majumdar & Ramchandran, ICIP 04) No need to deterministically track Y b and Y g at encoder. No need for multiple prediction loops complexity savings. Multiple side-informations at each decoder motion search at each decoder. Standards-compliant implementations possibly using the auxiliary channel setup. (Wang, Majumdar, & Ramchandran, ICASSP 05) Decoder Bad Decoder Good Encoder X Y g (good side-information) Y b (bad side-information) XgXg XbXb Rate = R Rate = R