International organisation for standardisation organisation internationale de normalisation

5MPEG-7 Visual

5.1MPEG-7 Visual related work in Shenzhen

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. All these documents were reviewed and discussed.

14907	Sangyoun Lee	Report of core experiment VCE-5
14909	Dae Il Yoon, Jie Jia, Wei Wu, Hae Kwang Kim, Weon Geun Oh	Video signature
14913	Hyeong-yong Jeon, Chi-jung Hwang, Weon-geun Oh	An Image Identifier Based on Feature Points for Complex Conditions
14926	Kota Iwamoto, Ryoma Oami	On the Modification Process of VCE-7
14927	Kota Iwamoto Ryoma Oami	Contribution to the Video Dataset for VCE-7
14956	Weon-Geun Oh, Ayoung Cho, Ik-Hwan Cho, Won-Keun Yang, Ju-Kyong Jin, Jun-Woo Lee, Dong-Seok Jeong	Concentric circle partition-based image signature
14960	Dong-Seok Jeong, Ju-Kyong Jin, Sang-Il Na, Dong-Jin Seo	Proposal on the Implementation of VCE-7 Analog VCR Recording & Recapture Modification
14964	Jens-Rainer Ohm	Responses received on Visual Signature Tools CfP
14977	Karol Wnukowicz	Cross verification result of Image Signature (VCE-6)
14983	Paul Brasnett, Miroslaw Bober	Proposed Improvements to Image Signature XM 31.0
14987	Weon-Geun Oh, Ik-Hwan Cho, Dong-Seok Jeong	Comments for the current geometrical modification of MPEG-7 VCE-6

On major work item has been the review of responses received on the Call for Proposals on Visual Signature Tools. Two proposals were received (as documented in M14964):

• 
  Proposal 1 (=14983): This is an improved version of the algorithm currently in the XM. Based on trace transform, projection to theta-axis, Fourier transform, log, difference between neighbored coefficients, binarization into 1008 bit. Scalable version: Splitting into subsets. Matching by Hamming distance. Correct detection at 0.05 ppm (1-recall) 99.59% in CE, 99.49% in CfP. No difference in descriptor, but different subset of 512 bits was used in these two cases. Proponent reports post experiments with 354 bits, which gave 99.65% correct detection.
  
• 
  Proposal 2 (=14956): Based on differences (radially 16+32 and in 36 angular directions) within circular neighborhoods. Depending on difference above or below a threshold, a binary description is extracted. Descriptor size is 392 bits, actually used size 354 bits, the remaining bits always zero, correct detection rate 99.65% under CfP conditions. Current version of descriptor is not scalable.
  

Regarding results, extraction and matching complexity as well as compactness of the description, both methods could be regarded as completely equivalent. This is true for the simple conditions as described in the context of the CfP. It was therefore decided to perform additional comparison considering the “light complex conditions” for rotation, translation and cropping as currently under investigation in VCE 6. Mutual crosschecks were performed by the proponents. The results of this additional testing under 10 ppm false matches were as follows (proposal #1 vs. #2):

• 
  Rotation 10%: 100% vs. 100%
  
• 
  Translation 10%: 0.98% vs. 0.19%
  
• 
  Cropping 90%: 57.5% vs. 0.9%
  

This indicates that proposal #1 has slight advantages as compared to proposal #2. Even though far from perfectly matching the cropped case, at least some of the images would be found using a comparably simple technique (more advanced technology for the entire “complex conditions” is currently being explored in VCE 6). The changes as made in the proposal are therefore promoted to WD. For further testing in VCE 6, concentration on only complex conditions will be made, since the simple conditions seems to work practically perfectly.

Progress was made in collecting test material for preparation of the CfP on video signatures (VCE 7). Further offers were brought only during the Shenzhen meeting. It will be decided after the next round of CE 7 during the January meeting whether the content available is sufficient to run the video signature CfP as planned (see below).

The following timeline is planned for the ongoing work on visual signatures:

• 
  Image Signature
  

• 
  PDAM: 2008/01
  
• 
  FPDAM: 2008/04
  
• 
  FDAM: 2008/10
  
• 
  AMD: 2009/01
  

• 
  Video Signature
  
  • 
    Collection of video test data: 2007.10.10
    
  • 
    Finalisation of experimental conditions / Final Call for Proposals: 2008.01.18
    
  • 
    Submission deadline: 2008.04.16 (by 23.59 Hours GMT)
    
  • 
    Evaluation of answers: 2008.04.26 – 05.02 (During the 84th MPEG meeting and the weekend before: proponents are strongly advised to present their proposals in person.)
    

• 
  PDAM: 2008/04
  
• 
  FPDAM: 2008/10
  
• 
  FDAM: 2009/04
  
• 
  AMD: 2009/07
  

Other work items:

• 
  VCE 5: Core experiment on using MPEG-7 face recognition technology is ongoing.
  
• 
  23000-3 Amd.1: One incompatibility in the software for PP MAF, regarding the BiM interface, was resolved in the meantime.
  
• 
  Video augmentation by metadata: This topic was treated in an AHG, which was not very active before the Shenzhen meeting, but is expected to bring more evidence aout the importance of this topic next time.