OggIssues

Problems resulting from design of Ogg

Seeking and Editing Problems

• jagged edges
• wide variance in location of cotemporal data
• impossible to reconstruct all granulepos values around holes
• granulepos / timeval mapping inconsistencies
• poorly sorted streams are rife
• impossible to efficiently seek with noncontinuous data

Synchronization

• no absolute clock (no presentation timestamps)
• no way to correct for clock skew between audio/video encoding

Other Niggles

• end-time ordering
• except when we have non-continuous data
  • Ordering isn't only an issue for non-continuous data. In theory, an idiot can fit up to ~30 minutes of Speex audio (silence) in a single page (or 4 minutes of actual speech).
• inefficient lacing values for video
• ad-hoc granulepos retrofitting for video, CMML
• seeking is hard
• pages, and libogg's behaviour when creating them

What use are...

• serial numbers?
• packet numbers?
• pages?
• checksums?
  • Useful for audio (preventing ear damage), but could be optional for video

Cleaner Abstractions

• We should not need to know the type of a stream if we are not decoding the stream
• granulepos interpretations
• headers
• seeking
• cutting
• Skeleton goes some way towards fixing this

Libogg issues

• Stupid decision for flushing pages
• Makes it generally easy to build broken files.

Proposed solutions

Short-term workarounds (Ogg1-compatible)

• Don't use partial packets unless absolutely necessary
• If absolutely necessary, don't share the pages with other packets
• Specify that pages should not contain more than X ms of data (let's say 250-500 ms)
• Put Theora keyframes alone on their page??

Successor to Ogg

• It should be called (Ogg2|Ogg3|Ogg++|OggNG|Ogh|Foo|Dumplings)
• The design should be done from desired capabilities and desired properties
• These capabilities and properties should come from AV experts, web-page designers, system administrators, and users

Desired Capabilities

• Simple seeking
• Cleanly cuttable
• Robust to errors
• Composable
• Supports arbitrary stream types

• Low bit cost
• Streamable
• Easy to chunk
• Low decode cost
• Supports multiple streams of each type

Untied We Stand

• Can cotemporal data be colocated?
• streams & bundles
• great for cutting
• OK for demultiplexing
• “should” cut down on bit overhead
• hugely simplifies seeking

Gimme a Hint

• Can we add seeking hints to the stream?
• these can be tiny and infrequent
• awesome for standalone files
• what do we do when streaming?
• hint correction packets?
• is this turtles all the way down?
• Would an up-front index be better?

Rebuttal

Devil's Advocate

• These problems aren't unsurmountable
• but we're only finding some of them now, and we've been working around others for years
• Nobody will adopt another container format
• Nobody cares about <insert hated feature here> anyway
• Even if we have Ogg2, we'll still be stuck having to support Ogg1 and broken files