encoding stereo images
DESCRIPTION
Encoding Stereo Images. Christopher Li, Idoia Ochoa and Nima Soltani. Outline. System overview Detailed encoder description Demonstration Results Extensions Conclusions. System Overview (Encoder). DWT. Quant. Arith Enc. L. DCT. Re-order. Arith Enc. u se ME. Motion Estimation. - PowerPoint PPT PresentationTRANSCRIPT
Title
Encoding Stereo ImagesChristopher Li, Idoia Ochoa and Nima Soltani 1OutlineSystem overviewDetailed encoder descriptionDemonstrationResultsExtensionsConclusions2System Overview (Encoder)
R
LDWTQuantArith EncDWTQuantMotion EstimationDCTRe-orderArith EncArith EncHuff Encresidualsshift vectorsuse ME3Left ImageDaubechies-4 wavelet decomposition 5 levels for luminance, 4 for chrominanceUniform quantization with adaptive levelsEach component meets its own fraction of MSEArithmetic coding on the quantized residualsFrequency tables are sent for each arithmetic coder
Better than Haar if we haveWaveletDaubechies-4 wavelet5 levels for luminance, 4 for chrominanceSubtract 127Zero-pad to 544x960
4Left Quantization
5Left QuantizationMotion Estimation Enable SignalHeuristically choose differential vs. separate encoding of right imageCalculate MSEY wavelet coeffsEncode differentiallyEncode separatelyYesNo
6Right ImageMotion Estimation Block
Mex files speed up the search significantly7Right ImageResidual codingImpose residuals of Cb and Cr to be 0Use remaining fraction of MSE for Y componentCompute DCT of blockReshape using zig-zag orderingReplace remaining zeros in block with end of block characterPerform arithmetic coding
8Right ImageShift vector codingOfflineFind joint statistics of the shift vectors over the training setConstruct Huffman tableDuring run-time, encode shift vectors using this Huffman table
9Right ImageSeparately codedSame method as left imageD4 wavelet, with 5 levels for Y, 4 for Cb, CrUniform quantization with variable stepArithmetic coding with frequencies sent
10Writing to FileUnique quantization values encoded in header bitsArithmetic codersEncode frequencies, output length of sequence and sequence itselfHuffman encoders Length of sequence and sequence itselfTables stored offline
11DecoderPerform all the steps of the encoder in reverseDecode left image using inverse DWTRead motion estimation flag for right imageIf enabled, decode shift vectors and residualsElse, decode using inverse DWT
12Demonstration
13ResultsImageBits/pixel11.732420.756630.180740.716450.930461.239071.8104ImageBits/pixel81.718890.6453101.7766110.9341120.5879132.1404142.383714Block size
Estimate disparity between left and right imageSmaller block size => more accurate estimate of shift vectors, but more shift vectors to encode\15ExtensionsUse intra-block coding for right imageExplore using different waveletsImplement embedded zero trees in CExplore run-length coding furtherApply uniform deadzone quantizers
16ConclusionsImportant trade-off between bits allocated to shift data and residual dataArithmetic coding outperforms HuffmanReshaping the DCT blocks allows us to use information, such as its size, to our advantageUniform quantizer is faster, simpler and has less overhead than Lloyd-max quantizersMEX files reduce runtime significantly!
17Thank you
Questions?18