The INESC Porto group has participated in the search task (automatic and interactive). Our approach combines high-level features (the 39 concepts of the LSCOM-Lite set) with low-level features. We use a large set of low-level features with the intention of analysing as many facets as possible of each shot. The aggregation of large feature sets can be time consuming as it needs to be done at query time. We have developed the BitMatrix indexing method to speed up the search process. For each shot, binary signatures in the form of bit sequences are obtained in an o -line process. At query time, the query bit signature is compared to each of the shots signatures. The automatic run performs above the median, in spite of not using any classi er or any other knowledge sources except the translation of the query into LSCOM-Lite concepts.
[1]
Philip S. Yu,et al.
The IGrid index: reversing the dimensionality curse for similarity indexing in high dimensional space
,
2000,
KDD '00.
[2]
G LoweDavid,et al.
Distinctive Image Features from Scale-Invariant Keypoints
,
2004
.
[3]
Paul M. B. Vitányi,et al.
The Google Similarity Distance
,
2004,
IEEE Transactions on Knowledge and Data Engineering.
[4]
Ichiro Fujinaga,et al.
jAudio: Additions and Improvements
,
2006,
ISMIR.
[5]
Hans-Jörg Schek,et al.
A Quantitative Analysis and Performance Study for Similarity-Search Methods in High-Dimensional Spaces
,
1998,
VLDB.
[6]
Cristina Ribeiro,et al.
Multidimensional Descriptor Indexing: Exploring the BitMatrix
,
2006,
CIVR.