Organisation internationale de normalisation

Matching

• 
  Most the of complexity for descriptor matching lies in local descriptors coordinates decoding, and it is due for the 64 bits integers division in the arithmetic decoder.
  
• 
  On ARMv7 architectures an additional complexity is observed due the 64 bits integer division software emulation  ( _divdi3 function) because ARMv7 is a 32 bits architecture.
  
• 
  Thus on ARMv7 architecture about 80% of the complexity of the pairwise matching pipeline is due to decode keypoints coordinates.
  

Retrieval

On ARMv7 architectures, most of complexity (70%) is related to the local descriptors matching, while the complexity related to the coordinate decoding is not relevant.

Extraction:

Around 1 second per image (fast version)

• 
  On ARMv7 architectures, the extraction time for APL_lowmemory and ALP_BFlog mode are about 1.6 times higher than ALP_fast an ALP_normal mode.
  
• 
  On ARMv7 architectures, most of extraction complexity arises from keypoints detection and extraction stages.
  

Concerns were expressed regarding the complexity of the coordinate decoding, which takes approximately 80% of the time in pairwise matching. It is recommended that the implementation is reviewed by the proponets to see what improvements can be made.

13.0.0.1.1.1.1.1.23m35072 A CDVS library with minimal dependencies [Massimo Balestri, Gianluca Francini, Skjalg Lepsoy, Attilio Fiandrotti, Massimo Mattelliano]

TM code was restructured to form a library, which simplified its use. New API was defined. Some functions and interfaces were redefined. Performance has been maintained (it is identical).

Improving the efficiency of the TM11 keypoint descriptor extractor. Reduced the complexity of the ALP descriptor extractor by 30%, resulting in the overall extraction time drop by 10%. Fully equivalent in terms of processing and performance (identical results).

Agreed to create a branch in TM to enable experiments and consider replacing the main TM trunk when crosschecks and anchors are confirmed (i.e. at MPEG 111 MPEG meeting). Namespace should be changed to mpeg7cdvs.

13.0.0.1.1.1.1.1.24m35337 CDVS: Crosscheck of M35072 [Karol Wnukowicz, Stavros Paschalakis]

Crosscheck OK.

13.0.0.1.1.1.1.1.25m35075 CDVS: Spatial resolution for robotic navigation [Pedro Porto Buarque de Gusmao, Gianluca Francini]

CDVS applied in a robotic navigation scenario, for robot odometry. Block size 3 and block size 1 were used. On the path length of 89m, the accuracy achieved for block size 1 was 4.7cm. (block size 3 gave 73 cm) Conclusion: CDVS can be used in odometry tasks, however a smaller block size may be required. Request to add smaller block size. This possibility of adding block size 1 and its implications on the specification will be analyzed by the group.

13.0.0.1.1.1.1.1.26m35076 A MATLAB interface to CDVS [Alessandro Bay, Massimo Balestri, Gianluca Francini]

Provides an advanced matlab interface to CDVS, based on TM software revision presented in m35072.

13.0.0.1.1.1.1.1.27m35131 Improvements to CDVS pairwise matching pipeline [Miroslaw Bober, Syed Husain, Stavros Paschalakis, Karol Wnukowicz]

(Document withdrawn.)

m35391 CDVS: Removal of the orientation parameter from feature selection

(Late contribution-no objections) Use of orientation for feature selection significantly adds to extraction complexity as SIFT descriptors have to be computed for all key-points detected. Only a limited subset of the local descriptors is subsequently used. This contribution proposed to remove the orientation parameter from the feature selection module. The poposed change has negligible impact on the retrieval and pairwise matching performance (basically maintained) but helps to significantly reduce extraction complexity.

Successfully crosschecked by m35395: DVS: Crosscheck of m35391

The group agreed to introduce this simplification into the text of DIS.