CS :: Fall 2003 MPEG Video (Part 2) Ketan Mayer-Patel
CS :: Fall 2003 Last Time Overall MPEG bitstream organization. I-Frames Examples of many encoding techniques: –Subsampling (chrominance planes) –Transform Coding (DCT, zig-zag) –Run-length Encoding (AC coeffs) –Predictive Encoding (DC coeffs) –Entropy Encoding (Huffman encoding) –Quantization (All coefficients)
CS :: Fall 2003 This Time P and B frames –Motion compensation. Search techniques The problem with error measurements –Skipped macroblocks Quantization control –Variable bitrate vs. Constant bitrate DCT Artifacts –Spider noise –Blockiness
CS :: Fall 2003 P-Frames Two types of macroblocks in P-Frames: –I-Macroblocks. Just like macroblocks in a I-Frame DC term is differentially encoded from DC predictor –DC predictor is simply last coded DC term. –Predictor reset at slice boundaries. –Encoded as DC size followed by that many bits. AC terms –RLE’d as (run,value) pairs. Huffman encoded. –P-Macroblocks
CS :: Fall 2003 P-Macroblocks Macroblock Address Increment (variable) Macroblock Type (1-6 bits) Q Scale (5 bits) Luminance BlocksU BlockV Block Motion Vector (variable) Block Pattern (3- 9 bits) Macroblock Type determines if Q Scale, Motion Vector, or Block Pattern exist. One or all of the blocks may be absent in a P-Macroblock.
CS :: Fall 2003 Address Increment Each macroblock has an address. –MB_WIDTH = width of luminance / 16 –MB_ROW = row # of upper left pixel / 16 –MB_COL = col. # of upper left pixel / 16 –MB_ADDR = MB_ROW * MB_WIDTH + MB_COL Decoder maintains PREV_MBADDR. –Set to -1 at beginning of picture. –Set to (SLICE_ROW*MB_WIDTH-1) at slice header. MB address increment added to PREV_MBADDR provides current macroblock address. –PREV_MBADDR set to current macroblock address.
CS :: Fall 2003 Address Increment Coding Address increment coded using Huffman code. –33 codes for values (1-33). 1 is smallest (1-bit) 33 is largest (11-bits) –1 code for ESCAPE ESCAPE means add 33 to address increment code that follows. ESCAPS can be chained allowing any positive value to be encoded as an address increment. This occurs for I-Frames as well.
CS :: Fall 2003 MB Type Huffman coded. –7 possible codes (1 - 6 bits) Determine the following: –Intra or non-intra. –Q scale specified or not. –Motion vector exists or not. –Block pattern exists or not. Not all combinations are possible. Not all possible combinations are feasible.
CS :: Fall 2003 Quantization Scale 5 bits. Zero is illegal. Encoded as 1-31 which results in q-scale values of (2-62). –Odd values impossible to encode. Decoder maintains current q-scale. –If not specified, current q-scale used. –If specified, current q-scale replaced.
CS :: Fall 2003 Motion Vector Two components: –Horizontal and vertical offsets. –Offset is from upper left pixel of macroblock. –Positive values indicate right and down. –Negative values indicate left and up. –Offsets are specified in half pixels. Motion vector is used to define a predictive base for the current macroblock from the reference picture.
CS :: Fall 2003 Motion Vector Illustrated P-FramePreviously Decoded I- or P- Frame Prediction base does not have to be macroblock aligned. If predictive base is half-pixel aligned, bilinear interpolation is used. Whatever luminance pixels are picked out, corresponding chrominance pixels used to form chrominance prediciton.
CS :: Fall 2003 Motion Vector Encoding If no motion vector is present, then motion vector is understood to be (0,0). Horiz. component followed by vertical. Decoder maintains motion vector predictor. –Set to 0,0 at beginning of picture or slice or whenever an I-macroblock is encountered. –Difference between predictor and value is Huffman encoded. Actually a bit more complicated than this.
CS :: Fall 2003 Predictive Base P-Macroblocks always specify a predictive base: –Either motion vector picks out an area, or –No motion vector implicitly implies 0,0 (i.e., predictive base is same macroblock in reference frame.)
CS :: Fall 2003 Block Pattern The goal of motion compensation is to find predictive base that matches most closely with macroblock. –If match is really good, then no appreciable difference will need to be encoded at all. Block pattern indicates which blocks have enough error to warrant coding. Absence of block pattern indicates no blocks needed coding.
CS :: Fall 2003 Block Difference between pixels in prediction and macroblock is encoded as block: –9-bit input values –Still produces 12-bit coefficients –Sometimes called error blocks.
CS :: Fall 2003 Error Block Encoding Different quantization matrix is used. –Default is “16” in all coefficient positions. –Error blocks have lots of high frequency info. –No good perceptual correlation between frequencies of error coding and artifacts. DC no longer specially treated. –No differential encoding from predictor. All terms are zig-zag RLE’d and then (run,value) pairs Huffman encoded.
CS :: Fall 2003 P-Frame Review Macroblocks are either I-macroblocks or P-macroblocks. I-macroblocks just like macroblocks in I-frame. P-macroblocks define predictive base and encode the difference.
CS :: Fall 2003 Skipped Macroblocks If P-macroblock has (0,0) motion vector and no appreciable difference to encode, then can be skipped altogether. Skipped macroblock detected when address increment for next coded macroblock is detected. First block and last block of slice must not be skipped. Last slice must include lower right macroblock.
CS :: Fall 2003 Decoder State Updates DC predictors are reset whenever a P-macroblock or skipped macroblock is encountered. Motion vector predictors reset whenever I-macroblock is encountered.
CS :: Fall 2003 B-Frames B-frames have 4 macroblock types: –I-macroblocks –P-macroblocks Predictive base specified from previous reference frame. –B-macroblocks Predictve base specified from subsequent reference frame. –Bi-macroblocks. Predictive base specified from both reference frames.
CS :: Fall 2003 Skipped Macroblocks Handled slightly differently than P-frames. Skipped macroblock implies: –Same macroblock type as last encoded macroblock (i.e., P-, B-, or Bi-). –Motion vectors same a previous encoded macroblock. Compare to (0,0) assumption in P-frame. Also means that predictors not reset. –Can’t skip macroblock following an I-macroblock. Other state changes as per P-frames.
CS :: Fall 2003 Motion Compensation Provides most of MPEG’s compression. Relies on temporal coherence. Finding a good motion vector essentially a search problem. Evaluating “goodness” of a motion vector can be a bit tricky. MC is what makes MPEG asymmetric. –Harder to encode than to decode.
CS :: Fall 2003 Exhaustive Search The most obvious and easiest solution. Encoding time related to size of search window. Although time consuming, also embarrassingly parallel.
CS :: Fall 2003 Logarithmic Search Evaluate the search window with an even sampling of motion vectors. Take best and reevaluate in region of the motion vector with denser sampling.
CS :: Fall 2003 Predictive Search Motion vectors differentially encoded for a reason. –Tend to be correlated from one macroblock to the next. Use previous macroblocks motion vector as centering point for search. Or, use motion vector from same block in previous frame as center of search. My research is looking at using depth and other spatial info to guide encoding.
CS :: Fall 2003 Error Measurements Regardless of search algorithm, need to determine which motion vector is best. Simple measures: –Mean Squared Error –Mean Absolute Error –Minimum Difference Variance Fundamental problem is no good correlation between any simple metric and perceptual quality.
CS :: Fall 2003 VBR vs. CBR Two ways to handle bitrate: –Variable Bit Rate (VBR) Allows compressed bitrate to vary –Constant Bit Rate (CBR) Bitrate constant over some averaging window. MPEG buffer model. –Optional (don’t have to use it). –Provides in the sequence header parameters to a buffer model that can describe bitrate behavior.
CS :: Fall 2003 VBR Q-scale adjustments In general, VBR used to maintain quality. Q scale is adjusted to provide maximum compression given quality limit. Need some metric for quality. –Same issue for judging perceptual quality crop up here. Common solution: q scale statically set for I-, P-, and B-frames. –A variation on this is differentiating among macroblock types.
CS :: Fall 2003 CBR Q-scale adjustments To achieve CBR, q-scale used to control bitrate. –Higher q-scale provides better compression at the expense of quality. –Lower q-scale provides better quality at the expense of compression. Algorithms for controlling how q-scale is adjusted can get pretty complicated. Common solution is to have target I, P, and B frame sizes and then adjust q-scale as macroblocks are encoded to hit the target.