Digital Media Dr. Jim Rowan ITEC 2110 Video Part 2 Digital Video Standards • Even though digital video COULD be much less complicated... it isn’t because... • Backward compatibility requirement – new equipment must create signals that can be handled by older equipment • Originally... TV signals (analog) needed to be converted to digital format Digital Video Standards... • Digital from NTSC and PAL are ANALOG standards – Each has a screen size and frame (or refresh) rate – Each define a number of lines on the screen (that can be easily used for one the Y dimension) – But what about the other dimension, the X?... – Each line is a continuous (analog) signal which has to be converted to digital... • How do you do that? – SAMPLE the analog data! – But directly sampling for each pixel results in a data stream of 20 Mbytes/ second... HUGE! Coping with Video Size • Aside from screen size and frame rate... • Consider human vision limitations – Use algebra to compute part of the signal – Chrominance sub-sampling • Compression - two versions – spatial – temporal Coping with Video Size • Aside from screen size and frame rate... • Consider human vision limitations – Use algebra to compute part of the signal – Chrominance sub-sampling • Compression - two versions – spatial – temporal Sampling analog Video • To reduce the data stream you can consider human vision again • Human eyes are less sensitive to color changes than luminance • Decision: Take fewer samples for color than luminance • Without sub-sampling... – for each pixel on the screen 4 things will have to be encoded • luminance, red, blue, green Sub-sampling & understanding human vision Designers realized that Green contributes the most to intensity, Red is next and Blue hardly contributes anything to luminance Based on this, it was decided to use a formula for luminance • Y = 0.2125R+0.7154G+0.0721B With this we only have 3 data elements to transmit 25% data reduction -Y (luminance) -Cb (blue chrominance) -Cr (red chrominance) Calculating the 4th color component • Known as the Y’CbCr model for CRTs Y = 0.2125R+0.7154G+0.0721B Solve for Cg (green): Y - 0.0721B - 0.2125R = 0.7154G 0.7154G = Y - 0.0721B - 0.2125R G = (Y - 0.0721B - 0.2125R) / 0.7145 There is a use for algebra! Coping with Video Size • Aside from screen size and frame rate... • Consider human vision limitations – Use algebra to compute part of the signal – Chrominance sub-sampling • Compression - two versions – spatial – temporal Chrominance sub-sampling • Humans can’t distinguish changes in color as well as they can distinguish luminance changes • http://en.wikipedia.org/wiki/Chroma_sub sampling • Of every 4 frames – store the luminance – only store a proportion of the color info Chrominance sub-sampling IRGB IRGB IRGB IRGB IRB I IRB I:R:B 4:4:4 I 4:2:2 CCIR 601 video sampling I I IRB IR IB I I I I I 4:1:1 NTSC DV I I 4:2:0 PAL DV notice the inconsistency? NTSC & PAL weirdness (sidebar) • NTSC & PAL – – – – Define different screen sizes Define different frame rates Both have the same aspect ratio of 4:3 BUT... they each are digitized (through sampling) to the same screen size • The result? – The pixels are not square – PAL is taller than it is wide – NTSC is wider than it is tall DV and MPEG • DV and its different forms: – MiniDV, DVCAM & DVPRO • http://en.wikipedia.org/wiki/DV#DVCAM • DVCAM & DVPRO – use the same compression algorithm (5:1) – use the same data stream (25Mbits) – use 4:2:2 sampling where DV uses 4:1:1 DV and MPEG • MPEG-1 originally meant for Video CD – http://en.wikipedia.org/wiki/Mpeg – never got very popular – developed into a family of standards • MPEG-4 rose from the ashes – http://en.wikipedia.org/wiki/Mpeg-4 – used for iTunes video A Computational Irony • Digital has been touted as a way to create exact copies while analog (VCR) cannot... – Analog VCR suffers from generational loss – Digital doesn’t suffer from generational loss • BUT only if you use video compression that is... LOSSLESS • AND... you guessed it, a lossless video compression technique is not used because the lossless ones don’t compress enough Lossless compression Original compression routine Exact duplicate Original compressed original decompress routine Lossy compression Original compression routine Changed Original compressed original decompress routine Changed Original 2 The Moral? • In production, if several people are working on the same bit of video, make sure that they all get uncompressed video to work with. • Only produce the compressed version after all the work is complete. Coping with Video Size • Aside from screen size and frame rate... • Consider human vision limitations – Use algebra to compute part of the signal – Chrominance sub-sampling • Compression - two versions – spatial – temporal Coping with Video Size • Spatial compression • Individual images can be compressed using the techniques discussed in the bitmapped section • Doesn’t result in very much compression for video • Doesn’t take into consideration the other frames that come before or after it Coping with Video Size • Aside from screen size and frame rate... • Consider human vision limitations – Use algebra to compute part of the signal – Chrominance sub-sampling • Compression - two versions – spatial – temporal Temporal Compression 1 • Use the Difference in two frames – naive approach can result in good compression – works well for a small amount of movement – A Tarantino film? not so much... Temporal Compression 2 • When an object moves – compute its trajectory – fill in the resulting exposed background vector – BUT there’s a problem... – why isn’t this an easy thing to do? Temporal Compression 2 • Bitmapped images do not have defined objects... that’s Vector graphics... • What to do? Temporal Compression 2 • Define blocks of 16 x 16 pixels – called a macroblock • Compute all possible movements of the block within a short range • Compute a vector to define that movement • Store the movement vectors • Compress the vectors More on Temporal Compression • Need some place to start from • Can be forward or backward prediction • Called KeyFrames – – – – pick a keyframe compute next image from that compute next image from that What happens when the scene completely changes? • Pick a new key frame... • But HOW? • Requires powerful AI Video Compression What does this? • Coder/Decoder - Codec – http://en.wikipedia.org/wiki/Video_codec • encodes and decodes video – can be symmetric • it takes as long to compress as decompress – can be asymmetric • it takes longer to compress or decompress than it does to decompress to compress A final worry... • We have been talking about making video smaller • There are a variety of techniques to do this • Which to choose? – It is a tradeoff between compression technique and its computational complexity Digital Media Dr. Jim Rowan ITEC 2110 Video Part 3 MPEG-4 • Designed for streams that contain video, still images, animation, textures 3-D models • Contains methods to divide scenes into arbitrarily shaped video objects • The idea is that each object has an optimal compression technique • BUT... MPEG-4 • Dividing a scene into arbitrarily shaped video objects is non-trivial – so they drop back to the rectangular object position • Quicktime and DivX use the rectangular video object idea • Uses forward interframe compression • Using the simpler technique reduces the computational complexity allowing it to be implemented on small devices like portable video players Other codecs • Cinepak, Intel Indeo & Sorenson • All use “vector” quantization – divides frame into rectangular blocks – these frames are called “vectors” but they don’t represent movement or direction • Codec uses a collection of these “vectors” – contains typical patterns seen in the frames • textures, patterns, sharp and soft edges – compares the “vectors” to the ones in the code book – if it is close, it uses the code book entry – (does this explain the patchwork painting of the screen when the digital signal goes bad?) “Vector” quantization • a frame contains indices into the code book • reconstructs the image from “vectors” in the code book – makes decompression very straight forward and efficient – this makes the implementation of a player very easy • What about the compression? “Vector” quantization • this is an asymmetric codec • decompression takes ~150 times longer than compression • Cinepak and Intel Indeo use temporal compression and simple differencing • Sorenson uses motion compensation similar to the MPEG-4 standard So... How do codecs vary? • compression and decompression complexity – affects the artifacts that are created – affects the time required to carry them out – affects the volume of the data stream created – affects the type and expense of the equipment used – affects whether or not it can be implemented in hardware of software Comparison Bear in mind that this comparison is not absolute and will vary from frame to frame but in general... • MPEG-4 – detail is good (at the sacrifice of speed) • DV – detail is good but the biggest • Sorenson – loss of detail (see pg 218-219) • Cinepak – loss of detail – smallest file A word about QuickTime • All standards so far have defined the data stream... not the file format • QT is the defacto standard design of a component-base architectural framework – allows plugins (components) to be developed by others • Every QT movie has a “time base” – records playback speed and current position relative to a time coordinate system to allow them to be replayed at the right speed on any system – If the playback speed of the device is not fast enough, QT drops frames keeping audio synchronization More about QuickTime Plugins make it flexible so that it can accommodate new file formats – comes with a standard set of plugins (components) • compressor components include – MPEG-4, Sorenson and Cinepak • movie controller interface provides uniformity • transcoder components to convert to other formats – supports true streaming and progressive download Questions?