Video Compression Demystified - Intro to MPEG, AVI, WMV
10 likes1,000 views
Video compression techniques exploit correlations in pixels over space and time, as well as limitations of human perception, to reduce file sizes. Key techniques include discrete cosine transform (DCT) to concentrate information, motion estimation and compensation to reduce temporal redundancy, and entropy coding like Huffman or arithmetic coding to remove statistical redundancy. Newer codecs like H.264 provide even greater compression by using smaller block sizes, quarter pixel motion estimation, multiple reference frames, and more efficient entropy coding.
Video Compression Demystified - Intro to MPEG, AVI, WMV
- 1. Video Compression Demystified Intro to MPEG, WMV, H.264, AVI
- 2. Why compression? Bandwidth constraint Space constraint I don’t know what else?
- 3. Basic idea Correlation (Entropy) Space, Time, Bits Human eye models Low-pass filter Brain ‘fills’ in for sharpness Frugality
- 4. Pixels and frames Each picture is a set of pixels in two dimensions, third being time. Finite sharpness, resolution issue. Each group of 8 pixels = block 2x2 blocks = Macroblock
- 5. Correlation in space Space is ‘gravitated’ towards being continuous. How do you exploit it? DCT inside a block AC-DC prediction across blocks
- 6. Correlation in Time The WHOLE image does not change. Do NOT repeat information. Motion estimation Locate where my marker is moving towards. Send only the coded difference. Bi-predict pictures.
- 7. Correlation in bits Run-level-last coding for zeroes. Subsequent bits still have correlation in them. Huffman coding (VLC) Arithmetic coding (CABAC)
- 8. Human eye Eye is not responsive to every detail Quantization Smoothening filters Eye is more responsive to brightness Less effort of chrominance Brain fills in crucial details.
- 9. DCT and quantization Random in Time domain is high freq. Spatial redundancy reduced to a few significant values. Zigzag coding + quantization Focus on lower freq and suppress high freq.
- 10. Motion Estimation Find the best match around – ME Diamond, Hierarchical, Logarithmic Generally over a block. Half-pel and quarter pel motion estimation using filters. Better match in sub-pel level.
- 11. RLL and Huffman Long sequence of zeroes helps reducing to smaller chunks of data. This chunk also has a probabilistic models. Exploit probability of symbols and be frugal in bits. Context adaptive arithmetic coding.
- 12. MPEG MPEG-1: DCT Pixel level ME AC-DC prediction P pictures Huffman coding
- 13. MPEG-4 DCT Quarter-pel ME B pictures AC-DC prediction Data partitioning Objects – layers, groups, streams
- 14. H.264 Integer transform on 4x4 block Quarter pel ME on 4x4 In-loop de-blocking Multiple reference pictures CABAC – 2 times frugal Bit-plane coding
- 15. AVI Audio Video Interleave format One frame of video with equivalent audio frame(s) Necessary information Display info Shit