International organisation for standardisation organisation internationale de normalisation
MPEG-7 4.1MPEG-7 related work in Busan
Download 3.13 Mb.
|
- Navigate this page:
- 4.1.1Review of Core Experiments and approval of the AHG recommendations.
- 4.1.1Review of new contributions and establishing Core Experiments
- 4.1.2New XM
- 4.1.4Photo Player MAF
- 4.1.5WD and error report for part 8
4MPEG-74.1MPEG-7 related work in BusanThe MPEG-7 breakout group was active during the whole week. Input documents as reviewed are listed in the table below.
Work performed: 4.1.1Review of Core Experiments and approval of the AHG recommendations.Core experiment VCE-1 evaluated description schemes, algorithms, and similarity measures applicable to the task of images/photo clustering into clusters depicting different situations/events. There is already one non-normative clustering algorithm in the XM, based solely on the visual descriptors, achieving recall=84% at precision=83%. Updated results for the algorithm using ‘visual semantic hints’ (M11965) were presented demonstrating good performance. The visual semantic hints of an image are used to weight the contribution from visual descriptors to reflect their perceptual importance. The weighting was further optimized with the EER performance around 92% (recall=94% at precision=90%). The algorithm was included into XM as a second non-normative tool for situation/view based image clustering. (Results cross-verified in M12014). Core experiment VCE-2 evaluated description schemes, algorithms, and similarity measures in the task of categorisation (labelling) of image/photo data. Relevant contributions include: M11966-CE report on category based photo clustering (VCE-2) and M12049-Results on Image Categorisation (VCE-2). The test dataset was now updated and allows multiple category labels for images. M11966 presented an extension of the original method from M11650 - it introduces 20 local concepts and uses them for global concept learning. Average performance (weighted by the number of ground truth images for each concept) is about 74% recall and 59% precision (66% average). M12049 performs pattern classification directly on the MPEG-7 visual descriptors extracted from each of the images (5 visual descriptors used) and employs k-Nearest Neighbour (kNN) classifier. Average performance (weighted by the number of ground truth images for each concept) is about 75% (for a mixed Corel/MPEG training dataset) and about 57% when training on Corel images and testing on MPEG images The video group agreed with the AHG group recommendation to continue this CE with the objecting to further improve the categorization performance and to obtain comparable results (i.e. the same testing scenario) for all techniques proposed. VCE-3 evaluated description schemes, algorithms, and similarity measures in the task of clustering and retrieval of images based on persons ID obtained from face descriptors. This CE had two parts. VCE3-Part 1 related to Person Identity based photo clustering. M11968 (Test dataset on face-based photo clustering) describes the images and GT used in experiment. The test dataset includes 1819 faces in 31 different situations. M11967 (CE report on face based photo clustering (VCE-3))- in the approach presented, the digital home photos are sorted by taken time and clustered by person identity first. The grouped photos are then indexed into the Database, which includes pre-registered faces. In the clustering experiment the error rate is 23.03% for the AFRD descriptor alone and 14.02% for combined AFRD combined with DC descriptor. In the indexing part, based on a subset of the dataset, the performance achieved with AFRD alone is 72.31% and 93.56% for the AFRD combined with DC. Video group agreed to continue this CE with expanded DB. VCE3-Part 2 focused on evaluation of MPEG 7 Face Recognition descriptor on Video Content (images extracted from digital video). M11924 (Preliminary results on VCE-3-Part 2) reported preliminary results of performance evaluation of FRD (Face Recognition Descriptor) and AFRD (Advanced Face Recognition Descriptor) on the KBS dataset, which consists of still images including faces captured from various KBS broadcasting videos (approx 5,650 faces images or 48 actors). The KBS dataset is expected to be released to MPEG-7 community shortly after the Busan meeting. The ANMRR results obtained are 0.67 for FRD and 0.51 for AFRD. Video group concluded that the performance is of existing MPEG-7 Face Recognition descriptors for video broadcast material (with significant pose variations and other distortions) is not satisfactory and starting a separate CE to investigate means of improving performance for such applications. 4.1.1Review of new contributions and establishing Core ExperimentsM12034 -Results of the use of MPEG-7 XM-based SCHEMA Reference System for segmentation algorithm evaluation – An interesting presentation showing the use of MPEG-7 technology within an image search and retrieval system build by the Schema project. The MPEG-7 visual descriptions are computed from image regions obtained from different algorithms for automated segmentation. M12035 Application of the MPEG-7 XM-based SCHEMA Reference System to art image retrieval – describes the application of this MPEG-7 XM-based system to the indexing and retrieval of 20000 annotated art images of the ALINARI [3] collection. M12109 contribution demonstrates that there may be some more robust Face Recognition technology (as compared to standardized MPEG-7 FRC and AFRD tools) available for generic video content (i.e. unconstraint view and significant lighting variations). As a result, a core experiment was set up to prepare evaluation databases and to investigate further how significant improvements can be achieved. Core experiment VCE-1 on Situation/view based image clustering has been completed. Three Core Experiments were defined: VCE-2 on Image Categorisation into classes, VCE-3 on Identity-based clustering, indexing and retrieval of images, and VCE-4 on Robust Face Recognition technology. See N7090 “Description of Core Experiments for MPEG-7 New Visual Extensions” for details. 4.1.2New XMThe text of MPEG-7 Visual XM was updated with the latest tools accepted as non-normative and version 24 was issued. 4.1.3DoC on ISO/IEC 15938-3:2002/PDAM 2 and FPDAM of perceptual shapeThe text of ISO/IEC 15938-3:2002/FPDAM2 containing 3D perceptual shape descriptor has been issued. We also requested a change to its name to “Perceptual 3D Shape Descriptor” to better reflect its content. The software is already included in the XM. A Disposition of Comments on ISO/IEC 15938-3:2002/PDAM 2 was also issued. 4.1.4Photo Player MAFSignificant work has been done to create the first working draft of the Photo Player MAF. There were joint discussions with Requirements, MDS and Systems on the possible MAF tools and structure. Several different options to carry MPEG-7 meta-data information in an ISO file format are proposed. An AHG was established to continue the work between meetings. 4.1.5WD and error report for part 8We issued a Working Draft WD2.0 of ISO/IEC TR 15938-8:2003/Amd.2 New Visual Extensions (N7092). 4.1.6Corrigendum item on ISO/IEC 15938-8 TRWe issued corrigendum on ISO/IEC 15938-8 TR fixing a problem with the definition of inverse DCT for the Colour Layout Descriptor. (N7091). Directory: sc29 -> open -> 29view 29view -> International organisation for standardisation organisation internationale de normalisation 29view -> International organisation for standardisation organisation internationale de normalisation Download 3.13 Mb. Share with your friends:
|