Glossary
The key terms and every abbreviation the course uses, what each means, and where it is explained.
Key terms
| Term | Meaning |
|---|---|
| arithmetic coding | Coding a whole sequence as one number inside an interval that each symbol narrows in proportion to its probability, so a symbol can cost a fraction of a bit.Explained on: Arithmetic coding |
| banding | Visible contours where a gradient should be smooth. They come from rounding steps that are large enough to see: too few levels, or too coarse a quantizer.Explained on: Bit depth and banding |
| bit depth | The number of bits in each sample, which sets how many integer codes it can take: 256 at 8 bits, 1,024 at 10.Explained on: Bit depth and banding |
| bitrate ladder | The list of picture sizes and bitrates a title is encoded at, from which a player picks.Explained on: Convex hull and per-title encoding |
| chroma | The two color-difference planes, Cb and Cr: how far blue and red sit from luma. At 8 bits, 128 is the neutral value of each, and a pixel with both at 128 is gray.Explained on: Pixels, luma and chroma |
| chroma subsampling | Storing the chroma planes at lower resolution than luma: half the width in 4:2:2, half the width and half the height in 4:2:0.Explained on: Chroma subsampling |
| coding standard | A specification of what a bitstream means and how a decoder rebuilds pictures from it, such as H.264. It does not say how an encoder makes its choices.Explained on: The hybrid codec loop |
| coefficient | One output of the transform: how much of one pattern the residual block contains.Explained on: The discrete cosine transform as a change of basis |
| context | The group a value is sorted into before it is coded, by what kind of value it is and what was coded nearby. Each context keeps its own probability estimate.Explained on: Context modeling |
| contrast sensitivity | One over the smallest contrast a viewer can detect. It depends on how fine the detail is.Explained on: What the eye misses: contrast sensitivity and masking |
| convex hull | In per-title encoding, the boundary of efficient trial encodes across picture sizes: those that no other encode matches or beats in quality with as few bits. A strict geometric convex hull can leave some of them out.Explained on: Convex hull and per-title encoding |
| deblocking filter | A filter that smooths across block edges where the step looks like a coding artifact and leaves edges that look like part of the picture.Explained on: Deblocking and boundary strength |
| dither | Noise added before rounding so that a gradient's steps are broken up: each pixel is noisier and the local average follows the gradient.Explained on: Bit depth and banding |
| drift | A mismatch between the reference pixels an encoder and a decoder predict from, which can carry into later predictions.Explained on: Why the encoder contains a decoder |
| early termination | Stopping a search before every candidate has been measured. The rule may be a judgment that the ones left are unlikely to win, or a bound that proves they cannot.Explained on: Mode decision and early termination |
| encoder implementation | A program that writes bitstreams for a coding standard and makes the choices the standard leaves open, such as x264 for H.264.Explained on: The hybrid codec loop |
| entropy | The average of minus log2 of the probability over a source's symbols, in bits. For symbols that do not depend on each other, no lossless code averages fewer bits per symbol.Explained on: Why some symbols cost less than a bit |
| film grain synthesis | Removing grain before encoding and sending a few parameters so the decoder can generate similar-looking grain and add it to its output.Explained on: Grain: retention versus synthesis |
| frame type | What the blocks of a frame may be predicted from: in an I frame, pixels of the same frame only; in a P frame, also a frame decoded earlier; in a B frame, also a combination of two such predictions.Explained on: Intra, predicted and bidirectional frames |
| in-loop filter | A filter applied to the reconstructed frame before it is stored as a reference, so that the encoder and every decoder must apply it identically.Explained on: Why filtering inside the loop helps prediction |
| irrelevance | Detail a viewer is unlikely to notice. Removing it is lossy; it is often called perceptual redundancy.Explained on: Four kinds of redundancy |
| keyframe | A frame the encoder offers as a place to begin decoding. In x264 an IDR frame, or with open groups of pictures an I frame carrying a recovery point.Explained on: Groups of pictures, open and closed |
| level | The integer that quantization assigns to a coefficient. Levels, not coefficients, are what gets coded.Explained on: Quantization and the quantization parameter (QP) |
| level (of a stream) | A set of limits on picture size, processing rate, bitrate and buffers that a stream declares it stays within. Not the level that quantization assigns to a coefficient.Explained on: Profiles, levels and tiers |
| lookahead | The frames an encoder examines, with rough cost estimates, before coding the first of them. It is where frame types and scene cuts are decided.Explained on: Scene cuts and lookahead |
| loop restoration | The last of AV1's in-loop filters: a Wiener or self-guided filter chosen per restoration unit by an encoder that compares against the source.Explained on: Directional enhancement and loop restoration in AV1 |
| luma | A weighted combination of the encoded red, green and blue values. It gives a gray version of the picture; it is not a measurement of light or of perceived brightness.Explained on: Pixels, luma and chroma |
| masking | Detail already in a picture making an added error harder to see. It depends on the detail and the error; a faint pattern can also make detection easier.Explained on: What the eye misses: contrast sensitivity and masking |
| mode decision | Choosing how to code a block from the candidates an encoder tries: skip, inter with some partition, or intra with some prediction.Explained on: Mode decision and early termination |
| parameter set | A unit that carries settings a decoder needs before it can decode pictures.Explained on: Bitstream structure |
| per-title encoding | Building the bitrate ladder for each title from trial encodes of that title, in place of one fixed table.Explained on: Convex hull and per-title encoding |
| plane | One of the grids of integers that make up a frame: in most video, one of luma and two of chroma.Explained on: Pixels, luma and chroma |
| pooling | Turning many scores, one per frame or per window, into one.Explained on: Test design pitfalls |
| prediction | The encoder's estimate of a block, formed from pixels that are already coded: neighbors in the same frame, or a block of an earlier frame.Explained on: Intra prediction |
| preset | A named bundle of encoder settings that trades encoding time for compression. Many of the settings limit how much the encoder searches; some turn coding tools on or off or change how quality is allocated.Explained on: Presets as search pruning |
| profile | The set of coding tools and sample formats a stream is allowed to use.Explained on: Profiles, levels and tiers |
| quality metric | A calculation that gives a number meant to track quality. The full-reference metrics in this course compare a decoded picture or video with its source.Explained on: Why quality metrics disagree |
| quantizer compression | How far the quantizer follows a frame's complexity. x264's qcomp of 1 gives equal quantizers, 0 nearly equal frame sizes.Explained on: Quantizer compression and frame-type offsets |
| rate control | Choosing the quantization parameter of every frame, and within frames, so that the stream meets a quality or size target.Explained on: Constant quantizer, constant rate factor, average bitrate and constant bitrate |
| rate factor | The constant a frame's complexity term is divided by to give its quantizer scale. Constant rate factor fixes it; average bitrate and a second pass search for it.Explained on: Constant quantizer, constant rate factor, average bitrate and constant bitrate |
| rate-distortion curve | Quality plotted against bits for several encodes of the same source by one encoder.Explained on: Rate-distortion curves and Bjøntegaard delta rate |
| reconstruction | The block the decoder ends up with: the prediction plus the residual rebuilt from the levels. It differs from the source by the quantization error.Explained on: Why the encoder contains a decoder |
| redundancy | Patterns or unequal probabilities that let data be represented more compactly without changing the decoded values.Explained on: Four kinds of redundancy |
| reference frame | An already coded frame, in its reconstructed form, that other frames are predicted from.Explained on: Multiple reference frames |
| residual | The source block minus its prediction. The transform and the quantizer work on the residual, not on the picture.Explained on: The hybrid codec loop |
| scene cut | A frame so different from the one before that predicting it costs about as much as coding it as intra.Explained on: Scene cuts and lookahead |
| search pattern | The set of positions a motion search tests around its current best match before it moves, such as a diamond or a hexagon.Explained on: Motion search patterns |
| start code | The byte pattern, zeros and then a one, that marks where a unit begins in a plain H.264 or HEVC stream.Explained on: Bitstream structure |
| symbol | One value handed to the entropy coder, such as a flag, a level or a prediction mode.Explained on: Why some symbols cost less than a bit |
| texture energy | How much a block varies, measured in x264 as the summed magnitudes of its Hadamard coefficients without the one that holds the average.Explained on: Psychovisual rate-distortion optimization |
| tier | In High Efficiency Video Coding and AV1, one of two bitrate ceilings for the same level. In the level tables read for this course the High tier exists from level 4 up.Explained on: Profiles, levels and tiers |
| variable-length code | A code that gives each symbol, or each group of symbols, a codeword of a whole number of bits, shorter for the likely ones.Explained on: Variable-length codes |
Abbreviations
| Term | Meaning | Used on |
|---|---|---|
| ABRaverage bitrate | A rate control mode that steers the whole stream toward a bitrate by feedback, without bounding any short stretch of it.Explained on: Constant quantizer, constant rate factor, average bitrate and constant bitrate | |
| ACnon-constant components | Every transform coefficient other than the constant component: the variation around the block's average.Explained on: The discrete cosine transform as a change of basis | |
| AQadaptive quantization | Varying the quantizer block by block within a frame, so that regions where errors are easier to see get finer quantization.Explained on: Adaptive quantization | |
| BD-rateBjøntegaard delta rate | The relative difference in rate between two encoders' rate-distortion curves at equal quality, as a percentage, averaged in log rate over the range of quality the curves share.Explained on: Rate-distortion curves and Bjøntegaard delta rate | |
| bSboundary strength | A number from 0 to 4 that H.264's deblocking filter gives each piece of a block edge, from how the two blocks were coded. It selects the filter and how far the filter may move a pixel; 0 means the edge is not filtered.Explained on: Deblocking and boundary strength | |
| CABACcontext-adaptive binary arithmetic coding | H.264's arithmetic coder: every value is turned into two-valued decisions. Most are coded with a probability estimate chosen by their context; some take an even-odds path and a few a near-certain one.Explained on: Arithmetic coding | |
| CAVLCcontext-adaptive variable-length coding | H.264's variable-length coder: values are written as whole codewords, with the table of codewords chosen by what was coded nearby.Explained on: Variable-length codes | |
| CBRconstant bitrate | In x264, one-pass average bitrate with a buffer model whose maximum rate is no higher than the bitrate, which makes the encoder use more of its budget. Frame sizes still vary; strict constant bitrate also pads the stream with filler.Explained on: Constant quantizer, constant rate factor, average bitrate and constant bitrate | |
| CDEFconstrained directional enhancement filter | An in-loop filter of AV1 that finds the direction of each 8x8 block and smooths along it, limiting how much a very different neighbor can pull a pixel.Explained on: Directional enhancement and loop restoration in AV1 | |
| CRFconstant rate factor | A rate control mode that sets an encoder rate factor, not a bitrate or a measured quality: each frame's quantizer follows its complexity around a level the setting fixes.Explained on: Constant quantizer, constant rate factor, average bitrate and constant bitrate | |
| CTUcoding tree unit | The largest block in High Efficiency Video Coding, up to 64x64 pixels, which is split into four repeatedly to form the blocks that are predicted.Explained on: Block partitioning | |
| CU-treecoding-unit tree | The x265 counterpart of macroblock-tree.Explained on: Macroblock-tree and coding-unit tree | |
| DCconstant component | The transform coefficient for the flat pattern. It is proportional to the block's average, not equal to it; the factor depends on how the transform is scaled.Explained on: The discrete cosine transform as a change of basis | |
| DCTdiscrete cosine transform | The transform that represents a block as a weighted sum of cosine patterns. In an encoder the block is usually a prediction residual, and the weights are the coefficients that get quantized.Explained on: The discrete cosine transform as a change of basis | |
| GOPgroup of pictures | The frames from one keyframe up to the next. Closed when nothing in it refers to a frame outside it; open when B frames just before the next keyframe may.Explained on: Groups of pictures, open and closed | |
| HEVCHigh Efficiency Video Coding | The video coding standard also known as H.265, the successor to H.264.Explained on: From H.264 to High Efficiency Video Coding, AV1 and Versatile Video Coding | |
| IDRinstantaneous decoding refresh | An intra frame after which, in decoding order, no frame is predicted from anything that came before it, so a decoder can start there with empty references.Explained on: Groups of pictures, open and closed | |
| MB-treemacroblock-tree | An x264 tool that lowers the quantization parameter of blocks its lookahead estimates to be useful references for later frames, relative to the frame's base level.Explained on: Macroblock-tree and coding-unit tree | |
| MSEmean squared error | The average of the squared differences between two pictures' pixels.Explained on: Why quality metrics disagree | |
| NALnetwork abstraction layer | The layer of H.264 and its successors that packs coded data into units, each with a header saying what it holds.Explained on: Bitstream structure | |
| OBUopen bitstream unit | AV1's unit of coded data: a header byte, an optional size, and a payload.Explained on: Bitstream structure | |
| PPSpicture parameter set | The unit that carries settings shared by many pictures, below the sequence parameter set.Explained on: Bitstream structure | |
| PSNRpeak signal-to-noise ratio | A quality score computed from mean squared error, on a log scale.Explained on: Why quality metrics disagree | |
| psy-RDpsychovisual rate-distortion optimization | In x264, a penalty added to a candidate's cost for how far the reconstruction's texture energy is from the source's, in either direction.Explained on: Psychovisual rate-distortion optimization | |
| QPquantization parameter | The number that sets the quantizer step size. Step size doubles every 6 steps of it.Explained on: Quantization and the quantization parameter (QP) | The discrete cosine transform as a change of basisQuantization and the quantization parameter (QP)Quantization matricesContext modelingDeblocking and boundary strengthSample adaptive offsetWhy filtering inside the loop helps predictionIntra, predicted and bidirectional framesHierarchical B-framesRate-distortion optimizationHow lambda follows the quantization parameterMode decision and early terminationConstant quantizer, constant rate factor, average bitrate and constant bitrateThe video buffering verifierTwo-pass encoding and complexity estimationQuantizer compression and frame-type offsetsBlock energyAdaptive quantizationAdaptive quantization in x264Macroblock-tree and coding-unit treePsychovisual rate-distortion optimizationGrain: retention versus synthesisRate-distortion curves and Bjøntegaard delta rateConvex hull and per-title encodingTest design pitfallsInteraction map: which settings fight and which need each other |
| RDrate-distortion | The trade between how many bits an encoding costs and how much error it leaves.Explained on: Rate-distortion optimization | |
| RDOrate-distortion optimization | Choosing between ways of coding a block by minimizing distortion plus lambda times bits.Explained on: Rate-distortion optimization | |
| RMSroot-mean-square | The square root of the average of squared values: a typical magnitude.Explained on: Block energy | |
| SADsum of absolute differences | A cheap distortion measure: the absolute pixel differences between two blocks, added up.Explained on: Distortion measures | |
| SAOsample adaptive offset | An in-loop filter of High Efficiency Video Coding that sorts reconstructed pixels into classes and adds one offset, measured by the encoder against the source, to every pixel of a class.Explained on: Sample adaptive offset | |
| SATDsum of absolute transformed differences | Like the sum of absolute differences, but taken after a Hadamard transform, so it tracks coding cost more closely.Explained on: Distortion measures | |
| SPSsequence parameter set | The unit that carries settings for a whole sequence: profile, level, picture size and which tools are on.Explained on: Bitstream structure | |
| SSDsum of squared differences | A distortion measure: the squared pixel differences between two blocks, added up.Explained on: Distortion measures | |
| SSIMstructural similarity | A quality score that compares local brightness, contrast and structure rather than raw error.Explained on: Why quality metrics disagree | |
| VBVvideo buffering verifier | The encoder's model of a decoder's input buffer: bits arrive at a maximum rate, each frame removes its size, and the level must stay at or above zero.Explained on: The video buffering verifier | |
| VMAFVideo Multi-Method Assessment Fusion | Netflix's quality metric: several per-frame measures turned into a score from 0 to 100 by a model trained on viewers' ratings.Explained on: Why quality metrics disagree | |
| VPSvideo parameter set | In High Efficiency Video Coding, a unit above the sequence parameter set.Explained on: Bitstream structure | |
| VVCVersatile Video Coding | The video format standardized after High Efficiency Video Coding, also called H.266.Explained on: From H.264 to High Efficiency Video Coding, AV1 and Versatile Video Coding |