The MPEG-DASH Media Presentation Description (MPD) is an XML document containing information about media segments, their relationships and information necessary to choose between them, and other metadata that may be needed by clients.
In this post, I describe the most important pieces of the MPD, starting from the top level (Periods) and going to the bottom (Segments). In a later post, I cover common informative metadata. Other topics that I might cover include MPD events, in-band events ('emsg'), and encryption (DRM).
For more information, refer to the latest version of ISO/IEC 23009-1, which ISO makes available for free.
Periods
Periods, contained in the top-level MPD element, describe a part of the content with a start time and duration. Multiple Periods can be used for scenes or chapters, or to separate ads from program content.
Adaptation Sets
Adaptation Sets contain a logical media stream or media streams that a user may want to only access a subset of. In the simplest case, a Period could have one Adaptation Set containing all audio and video for the content, but to reduce bandwidth, each stream can be split into a different Adaptation Set. A common case is to have one video Adaptation Set, and multiple audio Adaptation Sets (one for each supported language). Adaptation Sets can also contain subtitles or arbitrary metadata.
Adaptation Sets are usually chosen by the user, or by a user agent (web browser or TV) using the user's preferences (like their language or accessibility needs).
Representations
Representations allow an Adaptation Set to contain the same content encoded in different ways. In most cases, Representations will be provided in multiple screen sizes and bandwidths. This allows clients to request the highest quality content that they can play without waiting to buffer, without wasting bandwith on unneeded pixels (for example, a 720p TV doesn't need 1080p content). Representations can also be encoded with different codecs, allowing support for clients with different supported codecs (as occurs in browsers, with some supporting MPEG-4 AVC / h.264 and some supporting VP8), or to provide higher quality Representations to newer clients while still supporting legacy clients (providing both h.264 and h.265, for example). Multiple codecs can also be useful on battery-powered devices, where a device might chose an older codec because it has hardware support (lower battery usage), even if it has software support for a newer codec.
Representations are usually chosen automatically, but some players allow users to override the choices (especially the resolution). A user might choose to make their own representation choices if they don't want to waste bandwidth in a particular video (maybe they only care about the audio), or if they're willing to have the video stop and buffer in exchange for higher quality.
SubRepresentations
SubRepresentations contain information that only applies to one media stream in a Representation. For example, if a Representation contains both audio and video, it could have a SubRepresentation to give additional information which only applies to the audio. This additional information could be specific codecs, sampling rates, or embedded subtitles. SubRepresentations also provide information necessary to extract one stream from a multiplexed container, or to extract a lower quality version of a stream (like only I-frames, which is useful in fast-forward mode).
Media Segments
Media segments are the actual media files that the DASH client plays, generally by playing them back-to-back as if they were one continuous file (although things can get much more complicated when switching between representations). Formats will be covered in more detail by my post on profiles, but the two containers described by MPEG are the ISO Base Media File Format (ISOBMFF), which is similar to the MPEG-4 container format, and MPEG-TS. WebM in DASH is described in a document on the WebM project's wiki.
Media Segment locations can be described using BaseURL for a single-segment Representation, a list of segments (SegmentList) or a template (SegmentTemplate). Information that applies to all segments can be found in a SegmentBase. Segment start times and durations can be described with a SegmentTimeline (especially important for live streaming, so a client can quickly determine the latest segment). This information can also appear at higher levels in the MPD, in which case the information provided is the default unless overridden by information lower in the XML hierarchy. This is particularly useful with SegmentTemplate.
Segments can be in separate files (common for live streaming), or they can be byte ranges within a single file (common for static / "on-demand").
Index Segments
Index Segments come in two types: one Representation Index Segment for the entire Representation, or a Single Index Segment per Media Segment. A Representation Index Segment is always a separate file, but a Single Index Segment can be a byte range in the same file as the Media Segment.
Index Segments contain ISOBMFF 'sidx' boxes, with information about Media Segment durations (in both bytes and time), stream access point types, and optionally subsegment information in 'ssix' boxes (the same information, but within segments). In the case of a Representation Index Segment, the 'sidx' boxes come one after another, but they are preceded by an 'sidx' for the index segment itself.
Example
Before finishing, I'll include a commented example of an MPD, to show how these parts work together.
<?xml version="1.0"?>
<MPD xmlns="urn:mpeg:dash:schema:mpd:2011" profiles="urn:mpeg:dash:profile:full:2011"
minBufferTime="PT1.5S">
<!-- Ad -->
<Period duration="PT30S">
<BaseURL>ad/</BaseURL>
<!-- Everything in one Adaptation Set -->
<AdaptationSet mimeType="video/mp2t">
<!-- 720p Representation at 3.2 Mbps -->
<Representation id="720p" bandwidth="3200000" width="1280" height="720">
<!-- Just use one segment, since the ad is only 30 seconds long -->
<BaseURL>720p.ts</BaseURL>
<SegmentBase>
<RepresentationIndex sourceURL="720p.sidx"/>
</SegmentBase>
</Representation>
<!-- 1080p Representation at 6.8 Mbps -->
<Representation id="1080p" bandwidth="6800000" width="1920"
height="1080">
<BaseURL>1080p.ts</BaseURL>
<SegmentBase>
<RepresentationIndex sourceURL="1080p.sidx"/>
</SegmentBase>
</Representation>
</AdaptationSet>
</Period>
<!-- Normal Content -->
<Period duration="PT10M">
<BaseURL>main/</BaseURL>
<!-- Just the video -->
<AdaptationSet mimeType="video/mp2t">
<BaseURL>video/</BaseURL>
<!-- 720p Representation at 3.2 Mbps -->
<Representation id="720p" bandwidth="3200000" width="1280" height="720">
<BaseURL>720p/</BaseURL>
<!-- First, we'll just list all of the segments -->
<!-- Timescale is "ticks per second", so each segment is 1 minute
long -->
<SegmentList timescale="90000" duration="5400000">
<RepresentationIndex sourceURL="representation-index.sidx"/>
<SegmentURL media="segment-1.ts"/>
<SegmentURL media="segment-2.ts"/>
<SegmentURL media="segment-3.ts"/>
<SegmentURL media="segment-4.ts"/>
<SegmentURL media="segment-5.ts"/>
<SegmentURL media="segment-6.ts"/>
<SegmentURL media="segment-7.ts"/>
<SegmentURL media="segment-8.ts"/>
<SegmentURL media="segment-9.ts"/>
<SegmentURL media="segment-10.ts"/>
</SegmentList>
</Representation>
<!-- 1080p Representation at 6.8 Mbps -->
<Representation id="1080p" bandwidth="6800000" width="1920"
height="1080">
<BaseURL>1080/</BaseURL>
<!-- Since all of our segments have similar names, this time
we'll use a SegmentTemplate -->
<SegmentTemplate media="segment-$Number$.ts" timescale="90000">
<RepresentationIndex sourceURL="representation-index.sidx"/>
<!-- Let's add a SegmentTimeline so the client can easily see
how many segments there are -->
<SegmentTimeline>
<!-- r is the number of repeats _after_ the first one, so
this reads:
Starting from time 0, there are 10 (9 + 1) segments
with a duration of (5400000 / @timescale) seconds. -->
<S t="0" r="9" d="5400000"/>
</SegmentTimeline>
</SegmentTemplate>
</Representation>
</AdaptationSet>
<!-- Just the audio -->
<AdaptationSet mimeType="audio/mp2t">
<BaseURL>audio/</BaseURL>
<!-- We're just going to offer one audio representation, since audio
bandwidth isn't very important. -->
<Representation id="audio" bandwidth="128000">
<SegmentTemplate media="segment-$Number$.ts" timescale="90000">
<RepresentationIndex sourceURL="representation-index.sidx"/>
<SegmentTimeline>
<S t="0" r="9" d="5400000"/>
</SegmentTimeline>
</SegmentTemplate>
</Representation>
</AdaptationSet>
</Period>
</MPD>
Thanks to Kenrick for pointing out an error in this section regarding how @r is 0-based.
Conclusion
This should provide enough information to understand the structure of an MPD, and the general idea of how a basic DASH client works. Next time, I'll discuss additional metadata, which can be used to make a client much smarter, and provide a better user experience.