John HamannVickey Yeh

Compression of Stereo Images

OutlineBasic Algorithm

Color Compression

Motion Compensation

Conclusion

Basic AlgorithmDiscrete Cosine Transform

Uniform QuantizationNo Quantization Tables

Directly Store the DC Coefficients

Huffman Code the AC CoefficientsNo Run-Length Encoding

Zip the Final Binary Stream

Results

Color CompressionAlgorithm A

Mode 0: Encode Cr, Cb Mode 1: Encode Cr, Cb + Cr Mode 2: Encode Cr, Cb – Cr

Algorithm A VariantMode 0: Encode Cb, Cr Mode 1: Encode Cb, Cr + Cb Mode 2: Encode Cb, Cr – CbPerforms Markedly Worse

Algorithm A

Algorithm A

About 1% Improvement

Algorithm B• Estimate Cb, Cr based on Y

• Cb = acb Y + bcb

• Cr = acr Y + bcr

• For each block• Intra-mode• Code acb , acr , bcb , bcb

• For small blocks,• Too expensive

• For large blocks,• Too much MSE

Motion Compensation• Search for Nearest Match

• Make Decision Based on Lowest MSE

• Give Bonus to Certain Intervals to Cause Clustering of Motion Vectors

• Encode the Residuals Using DCT as Before

Results

ConclusionHuffman Coding then Zip outperforms

Run-Length Encoding then Huffman Coding

Color Compression Using Algorithm A3 Modes of Operation {Cb, Cb+Cr, Cb-Cr}Provides slight gainsLimited by nature of Cb and Cr

Motion CompensationDid not prove beneficial

Questions