(typo - duplicate field)
|Line 3:||Line 3:|
= Ogg Skeleton 3.3 with Keyframe Index =
= Ogg Skeleton 3.3 with Keyframe Index =
'''This specification is and Ogg . '''
'''This specification is
== Overview ==
== Overview ==
Latest revision as of 23:22, 22 November 2010
Ogg Skeleton 3.3 with Keyframe Index
This specification is obsolete and has been superseded by Ogg Skeleton 4. Use that instead!
Seeking in an Ogg file is typically implemented as a bisection search over the pages in the file. The Ogg physical bitstream is bisected and the next Ogg page's end-time is extracted. The bisection continues until it reaches an Ogg page with an end-time close enough to the seek target time. However in media containing streams which have keyframes and interframes, such as Theora streams, your bisection search won't necessarily terminate at a keyframe. Thus if you begin decoding after your first bisection terminates, you're likely to only get partial incomplete frames, with "visual artifacts", until you decode up to the next keyframe. So to eliminate these visual artifacts, after the first bisection terminates, you must extract the keyframe's timestamp from the last Theora page's granulepos, and seek again back to the start of the keyframe and decode forward until you reach the frame at the seek target.
This is further complicated by the fact that packets often span multiple Ogg pages, and that Ogg pages from different streams can be interleaved between spanning packets.
The bisection method above works fine for seeking in local files, but for seeking in files served over the Internet via HTTP, each bisection or non sequential read can trigger a new HTTP request, which can have very high latency, making seeking very slow.
Seeking with an index
The Skeleton 3.3 bitstream attempts to alleviate this problem, by providing an index of periodic keyframes for every content stream in an Ogg segment. Note that the Skeleton 3.3 track only holds data for the segment or "link" in which it resides. So if two Ogg files are concatenated together ("chained"), the Skeleton 3.3's keyframe indexes in the first Ogg segment (the first "link" in the "chain") do not contain information about the keyframes in the second Ogg segment (the second link in the chain).
Each content track has a separate index, which is stored in its own packet in the Skeleton 3.3 track. The index for streams without the concept of a keyframe, such as Vorbis streams, can instead record the time position at periodic intervals, which achieves the same result. When this document refers to keyframes, it also implicitly refers to these independent periodic samples from keyframe-less streams.
All the Skeleton 3.3 track's pages appear in the header pages of the Ogg segment. This means the all the keyframe indexes are immediately available once the header packets have been read when playing the media over a network connection.
For every content stream in an Ogg segment, the Ogg index bitstream provides seek algorithms with an ordered table of "key points". A key point is intrinsically associated with exactly one stream, and stores the offset of the page on which it starts, o, as well as the presentation time of the keyframe t, as a fraction of seconds. This specifies that in order to render the stream at presentation time t, the last page which lies before all information required to render the keyframe at presentation time t begins exactly at byte offset o, as offset from the beginning of the Ogg segment. The offset is exactly the first byte of the page, so if you seek to a keypoint's offset and don't find the beginning of a page there, you can assume that the Ogg segment has been modified since the index was constructed, and that the index is now invalid and should not be used. The time t is the keyframe's presentation time corresponding to the granulepos, and is represented as a fraction in seconds. Note that if a stream requires any preroll, this will be accounted for in the time stored in the keypoint.
The Skeleton 3.3 track contains one index for each content stream in the file. To seek in an Ogg file which contains keyframe indexes, first construct the set which contains every active streams' last keypoint which has time less than or equal to the seek target time. Then from that set of key points, select the key point with the smallest byte offset. You then verify that there's a page found at exactly that offset, and if so, you can begin decoding. If the first keyframe you encounter has a time equal to that stored in the keypoint, you have made the optimal seek, and can safely continue to decode up to the seek target time. You are guaranteed to pass keyframes on all streams with time less than or equal to your seek target time while decoding up to the seek target. However if the first keyframe you encounter after decoding does not have the same presentation time as is stored in the keypoint, you then the index is invalid (possibly the file has been changed without updating the index) and you must either fallback to a bisection search, or keep decoding if you've landed "close enough" to the seek target.
Be aware that you cannot assume that any or all Ogg files will contain keyframe indexes, so when implementing Ogg seeking, you must gracefully fall-back to a bisection search or other seek algorithm when the index is not present, or when it is invalid.
The Skeleton 3.3 BOS packet also stores meta data about the segment in which it resides. It stores the timestamps of the first and last samples in the segment. This also allows you to determine the duration of the indexed Ogg media without having to decode the start and end of the Ogg segment to calculate the difference (which is the duration).
The Skeleton 3.3 BOS packet also contains the length of the indexed segment in bytes. This is so that if the seek target is outside of the indexed range, you can immediately move to the next/previous segment and either seek using that segment's index, or narrow the bisection window if that segment has no index. You can also use the segement length to verify if the index is valid. If the contents of the segment have changed, it's highly likely that the length of the segment has changed as well. When you load the segment's header pages, you should check the length of the physical segment, and if it doesn't match that stored in the Skeleton header packet, you know the index is out of date and not safe to use.
The Skeleton 3.3 BOS packet also contains the offset of the first non header page in the Ogg segment. This means that if you wish to delay loading of an index for whatever reason, you can skip forward to that offset, and start decoding from that offset forwards.
When using the index to seek, you must verify that the index is still correct. You can consider the index invalid if any of the following are true:
- The segment length stored in the Skeleton BOS packet doesn't match the length of the physical segment, or
- after a seek to a keypoint's offset, you don't land exactly on a page boundary, or
- the first keyframe decoded after seeking to a keypoint's offset doesn't have the same presentation time as stored in the index.
You should also always check the Skeleton version header field to ensure your decoder correctly knows how to parse the Skeleton track.
Be aware that a keyframe index may not index all keyframes in the Ogg segment, it may only index periodic keyframes instead.
Unless otherwise specified, all integers and fields in the bitstream are encoded with the least significant bit coming first in each byte. Integers and fields comprising of more than one byte are encoded least significant byte first (i.e. little endian byte order).
The Skeleton 3.3 track is intended to be backwards compatible with the Skeleton 3.0 specification, available at http://www.xiph.org/ogg/doc/skeleton.html . Unless specified differently here, it is safe to assume that anything specified for a Skeleton 3.0 track holds for a Skeleton 3.3 track.
As per the Skeleton 3.0 track, a segment containing a Skeleton 3.3 track must begin with a Skeleton 3.3 fishead BOS packet on a page by itself, with the following format:
- Identifier: 8 bytes, "fishead\0".
- Version major: 2 Byte unsigned integer denoting the major version (3)
- Version minor: 2 Byte unsigned integer denoting the minor version (1)
- Presentationtime numerator: 8 Byte signed integer
- Presentationtime denominator: 8 Byte signed integer
- Basetime numerator: 8 Byte signed integer
- Basetime denominator: 8 Byte signed integer
- UTC [ISO8601]: a 20 Byte string containing a UTC time
- [NEW] First-sample-time numerator: 8 byte signed integer representing the numerator for the presentation time of the first sample in the media. Note that samples between the first-sample-time and the Presentationtime are supposed to be skipped during playback.
- [NEW] First-sample-time denominator: 8 byte signed integer, with value 0 if the timestamp is unknown. Decoders should always ensure that the denominator is not 0 before using it as a divisor!
- [NEW] Last-sample-time numerator: 8 byte signed integer representing the end time of the last sample in the segment.
- [NEW] Last-sample-time denominator: 8 byte signed integer, with value 0 if the timestamp is unknown. Decoders should always ensure that the denominator is not 0 before using it as a divisor!
- [NEW] The length of the segment, in bytes: 8 byte unsigned integer, 0 if unknown.
- [NEW] The offset of the first non-header page, in bytes: 8 byte unsigned integer, 0 if unknown.
The first-sample-time and last-sample-time are rational numbers, in units of seconds. If the denominator is 0 for the first-sample-time or the last-sample-time, then that value was unable to be determined at indexing time, and is unknown. The duration of the Ogg segment can be calculated by subtracting the first-sample-time from the last-sample-time.
In Skeleton 3.3 the "fisbone" packets remain unchanged from Skeleton 3.0, and will still follow after the other streams' BOS pages and secondary header pages.
Before the Skeleton EOS page in the segment header pages come the Skeleton 3.3 keyframe index packets. There should be one index packet for each content stream in the Ogg segment, but index packets are not required for a Skeleton 3.3 track to be considered valid. Each keypoint in the index is stored in a "keypoint", which in turn stores an offset, checksum, and timestamp. In order to save space, the offsets and timestamps are stored as deltas, and then variable byte-encoded. The offset and timestamp deltas store the difference between the keypoint's offset and timestamp from the previous keypoint's offset and timestamp. So to calculate the page offset of a keypoint you must sum the offset deltas of up to and including the keypoint in the index.
The variable byte encoded integers are encoded using 7 bits per byte to store the integer's bits, and the high bit is set in the last byte used to encode the integer. The bits and bytes are in little endian byte order. For example, the integer 7843, or 0001 1110 1010 0011 in binary, would be stored as two bytes: 0xBD 0x23, or 1011 1101 0010 0011 in binary.
Each Skeleton 3.3 keyframe index packet contains the following:
- Identifier 6 bytes: "index\0"
- The serialno of the stream this index applies to, as a 4 byte field.
- The number of keypoints in this index packet, 'n' as a 8 byte unsigned integer. This can be 0.
- The keypoint presentation time denominator, as an 8 byte signed integer.
- 'n' key points, each of which contain, in the following order:
- the keyframe's page's byte offset delta, as a variable byte encoded integer. This is the number of bytes that this keypoint is after the preceeding keypoint's offset, or from the start of the segment if this is the first keypoint. The keypoint's page start is therefore the sum of the byte-offset-deltas of all the keypoints which come before it.
- the presentation time numerator delta, of the first key frame which starts on the page at the keypoint's offset, as a variable byte encoded integer. This is the difference from the previous keypoint's timestamp numerator. The keypoint's timestamp numerator is therefore the sum of all the timestamp numerator deltas up to and including the keypoint's. Divide the timestamp numerator sum by the timestamp denominator stored earlier in the index packet to determine the presentation time of the keyframe in seconds.
Note that a keypoint always represents the first key frame on a page. If an Ogg page contains two or more keyframes, the index's key point *must* refer to the first keyframe on that page, not any subsequent keyframes on that page.
The key points are stored in increasing order by offset (and thus by presentation time as well).
The byte offsets stored in keypoints are relative to the start of the Ogg bitstream segment. So if you have a physical Ogg bitstream made up of two chained Oggs, the offsets in the second Ogg segment's bitstream's index are relative to the beginning of the second Ogg in the chain, not the first. Also note that if a physical Ogg bitstream is made up of chained Oggs, the presence of an index in one segment does not imply that there will be an index in any other segment.
The exact number of keyframes used to construct key points in the index is up to the indexer, but to limit the index size, we recommend including at most one key point per every 64KB of data, or every 2000ms, whichever is least frequent.
As per the Skeleton 3.0 track, the last packet in the Skeleton 3.3 track is an empty EOS packet.
For a prototype indexer, see OggIndex. Also included there is a program OggIndexValid, which can verify that Theora and Vorbis indexes are valid. If you're implementing your own indexer, or going to be modifying existing indexes, always verify that your modified indexes are valid as per OggIndexValid!
Recent ffmpeg2theora nightlies will also include a keyframe index in the Skeleton 3.3 track if you specify the command line option --seek-index.
To see how indexes improves network seeking performance, you can download a development version of Firefox which can take advantage of indexes here:
If you already have a Firefox instance running, you'll need to either close your running Firefox instance before starting the index-capable Firefox, or start the index-capable Firefox with the --no-remote command line parameter.
To compare the network performance of indexed versus non-indexed seeking, point the index-capable Firefox here: