SB task futures:

	Try keyframe selection (how to evaluate? how to choose among?)

	Try SB task again on new data - languages, sources, etc

	Try detection of low-level features like camera motion

	Try detection of various production artifacts

Feature extraction:

	Questions about use of (M)AP and assumption that not-judged =
	not-relevant

	Drop ranking, evaluate a sample of complete submissions for
	frequent features

	Provide complete feature judgments for use in search

	Keep relatively rare features since they occur in news

	Keep a few very frequent features

	Make sure test & development data have close to same encoding

	Increase the number of features greatly (use most frequent
	from common annotation?)

Story segmentation:

	Find stories around / start of enclosing story for search result

	Add topics which request stories (but are visual)

	Do video topic tracking (how to evaluate?) - see "find
	stories..." above

	What will "story" mean in the data for 2005?

	Repeat task with other (news) sources (use more recent TDT
	annotations?)

Search:

	Alternate search scenarios: novice searching public collection
	(via Web?); ...?

	Provide standard topic categorization for optional use

	Include identification of commerical shots with master shot ref

	Use master shot reference again to define units of feature extraction,
	retrieval? Is there an alternative?

	Will ASR output be provided to participants?

	Avoid search topic targets in commercials; adjust/correct measures; ...


New data:

	In selection and use, consider annotation provided


New tasks?:

	Add a task designed around "near duplicate detection"  - first
	step toward topic tracking?


	Add a pilot task dealing with a dataset where there is little
	or no text that can be OCR'ed from the frames and where search
	cannot leverage most of its performance from ASR. E.g., raw
	clips, instructional videos?, automatic indexing of conference
	captures, single camera/shot coverage of sports matches
	(e.g. for sports coaches illustrating their peptalk at
	breaktime)), single camera/shot coverage of e.g. operating
	rooms (to produce an indexed video, suitable for training),
	surveillance data: indexing behaviour of interest,
	summarization. Tasks could be defined related to getting an
	understanding of what happens in the video in terms of events,
	in order to build a semantic index into the video. This could
	be built up across a cycle of e.g. three years.

	Look for a "smaller-world" data set and task. If features can't
	be used there, no need to bother trying in the wider worlds
	of news etc? (How about game video of some sports team....

Venue?  

	ACM MM 05 (Singapore) (prohibitive travel costs for some?)

	NIST (Gaithersburg, MD, USA with TREC 05)

	CIVR 06 (Tempe, AZ, USA) -> 18mo.cycle (loss of momentum?; loss of summer for work)