International organisation for standardisation

4MPEG-7 Visual

4.1MPEG-7 Visual related work in Archamps

The MPEG-7 breakout group was active during the whole week. Input documents related to the Visual part in 15938-3 are listed in the table below.

m15347	CE report for VCE-5	Sangki- Kim Hyobin Lee Sangyoun Lee
m15358	Clarification of Matching in Image/Video Signature	Weon-Geun Oh Won-Keun Yang AYoung Cho Dong-Seok Jeong
m15383	Suggestion of Common Output Format for VCE-7 Experiment	Kota Iwamoto Ryoma Oami
m15434	Cross verification results of experiments for Image Signature (VCE-6)	Karol Wnukowicz
m15449	Proposal on Image Signature for Complex Conditions Combination scale-adapted Harris and scale-adapted Laplacian feature detector. Select 32 strongest feature points, represented by 60 bits each in trace transform, to represent the signature. 4-stage matching: Preselection of 24 features based on Hamming distance, check geometric constraints, establish second hypothesis based on 32 features (including geometric check). Highly improved performance for complex conditions (except heavy cropping and translation). Matching speed 110000 images/second according to proponents; others claim that the software they were given is only 4400 images/second. Clarify this offline. If clarification is successful and also current matching procedure is sufficiently fast, adopt this for XM. Perform more sanity check until next meeting, and include into Study of FPDAM by July when found sufficiently mature. Note: Matching is not normative, therefore matching speed should not be used as the most relevant criterion.	Paul Brasnett Miroslaw Bober
m15454	Video content for VCE-7	Paul Brasnett Modupe Gamu Miroslaw Bober
m15463	Method for Automatic modification software using various video peroperties for VCE-7	Ju-Kyong Jin Sang-IL Na Jun-Woo Lee Dong-Seok Jeong

Amendment 3 of MPEG-7 Visual (Image Signature Descriptors) was progressed to FPDAM as planned. Except for editorial improvements no changes were made relative to the PDAM (technology based on trace transform, which shows good results for the cases of finding the same images with moderate changes). In M15449, new results from visual core experiment 6 (VCE-6) are reported, where the extraction of the descriptor is additionally applied to local feature points (based on Harris & Laplacian corner/edge detectors). The results indicate that this method is also capable to meet the requirements of medium to heavy modifications (with least good – but significantly improved as compared to previous reports – results still in cases of heavy cropping and translation). It was decided to adopt this method to the XM with the potential to promote it into a Study of FPDAM by the next meeting, provided that no further deficiencies are found in the process of XM validation (continuing in VCE-6).

For Video Signatures (preparation of CfP, software for systematic modifications etc. in VCE-7), the expectations of having sufficient material available (approx. 200 hours of video content required, varied content such as: sports, news, film, soap opera, variety of others, approx. 4 000 longer clips and 24 000 shorter clips) have not been realized; in fact approximately 70 hours have been collected (all with copyright of using it in the context of MPEG-7 development). It was nevertheless decided to launch a preliminary CfP from the Archamps meeting, which will be updated by July after more content would be available; in the event that it proves difficult to acquire more content, it will be necessary to tweak the testing conditions appropriately in order to still fulfill the expected range of quality (50 million comparisons need to be performed for the envisaged successful hits vs. false alarm rate). Pre-registration for the CfP, as well as distribution of the available content to prospective proponents can immediately be started to avoid further delays. Nevertheless, due to the large amount of material, the logistic effort in the context of the VS CfP is large, such that the following timeline appears currently realistic:

• 
  Final CfP – 08/07
  
• 
  Responses due – 09/01
  
• 
  PDAM – 09/04
  
• 
  FPDAM – 09/10
  
• 
  FDAM – 10/04
  

The previous VCE-5 (face recognition with infrared images) has been discontinued. Apparently, during the lifetime of this experiment the purpose was changed from purely exploring the usability of existing MPEG-7 face description technology for this application case into a more generic investigation of various methods. It will be necessary to better understand the purpose of doing this, and whether this is a relevant work topic for MPEG, also considering the fact that more prominent development in face recognition technology may be performed elsewhere. Interested parties are invited by resolution to bring more evidence on this by the next meeting.

ISO/IEC 23000-3 / FPDAM 1 Photo Player MAF Reference Software (N9222 of July 2007 meeting) was finalized for release. All bugs were removed, and it was identified that the remaining mismatch issue is clearly due to a bug in the BiM reference software. Work on conformance testing could now finally be started, but was not yet possible to find volunteers on this yet. Therefore, an urgent resolution was issued announcing that the entire PP MAF specification may be removed unless the work on conformance is started.
New overview documents (on existing Visual Description Tools and the new Visual Signature Tools) were produced for the purpose of publication on the MPEG technologies web page.