Multi-tiered image clustering by event
First Claim
Patent Images
1. A method for classifying a sequence of digital records into events based upon feature values associated with each of said digital records, said method comprising using a processor to perform the steps of:
- determining feature differences between consecutive said digital records;
ranking said feature differences to provide an unclustered feature difference set;
computing a sequence of three or more mutually exclusive feature difference clusters, each feature difference cluster corresponding to a different probability of representing separations between events, wherein at least one boundary between feature difference clusters is constrained to fall within a specified range, the feature difference clusters being computed using a 2-means event clustering algorithm to determine at least one of said feature difference clusters and using a different clustering algorithm based upon a variance metric to determine at least one other of said feature difference clusters; and
classifying the sequence of digital records into events using the feature differences that fall into one or more of the feature difference clusters.
5 Assignments
0 Petitions
Accused Products
Abstract
In a method for classifying a sequence of records into events based upon feature values, such as time and/or location, associated with each of the records, feature differences between consecutive records are determined. The feature differences are ranked. A sequence of three or more clusters of feature differences is computed. The clusters are arranged in decreasing order of relative likelihood of respective feature differences representing separations between events. The records can be inclusive of images.
-
Citations
26 Claims
-
1. A method for classifying a sequence of digital records into events based upon feature values associated with each of said digital records, said method comprising using a processor to perform the steps of:
-
determining feature differences between consecutive said digital records; ranking said feature differences to provide an unclustered feature difference set; computing a sequence of three or more mutually exclusive feature difference clusters, each feature difference cluster corresponding to a different probability of representing separations between events, wherein at least one boundary between feature difference clusters is constrained to fall within a specified range, the feature difference clusters being computed using a 2-means event clustering algorithm to determine at least one of said feature difference clusters and using a different clustering algorithm based upon a variance metric to determine at least one other of said feature difference clusters; and classifying the sequence of digital records into events using the feature differences that fall into one or more of the feature difference clusters. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8)
-
-
9. A method for classifying a sequence of digital records into events based upon feature values associated with each of said digital records, said method comprising using a processor to perform the steps of:
-
determining differences between the feature values of consecutive said records to provide feature differences; ranking said feature differences to provide an unclustered feature difference set; sequentially partitioning a plurality of feature difference clusters of said feature differences from said unclustered feature difference set to form at least three feature difference clusters, said feature difference clusters being mutually exclusive, said partitioning being in decreasing order of relative likelihood of respective said feature differences representing separations between events, wherein at least one partition boundary between feature difference clusters is constrained to fall within a specified range, said partitioning comprising using a 2-means event clustering algorithm to partition at least one of said feature difference clusters and using a difference clustering algorithm based upon a variance metric to partition at least one other of said feature difference clusters; and classifying the sequence of digital records into events using the feature differences that fall into one or more of the feature difference clusters. - View Dependent Claims (10, 11, 12, 13, 14, 15, 16, 17, 18, 19)
-
-
20. A method for classifying a database of digital records into events, said digital records each having an associated feature value, said features values being ordinal and having a dimensionality of one or more, said records being ranked in order of respective said feature values, said method comprising using a processor to perform the steps of:
-
determining feature value differences between consecutive said digital records; ranking said feature value differences to provide an unclustered feature difference set; calculating an boundary from said feature value differences of said unclustered feature difference set, said boundary defining a feature difference cluster of said differences of said unclustered feature difference set relatively more likely to represent separations between events, said boundary redefining said unclustered feature difference set to exclude said feature difference cluster, said boundary being constrained to fall within a specified range; repeating said calculating at least once to form additional feature difference clusters; and classifying the sequence of digital records into events using the feature differences that fall into one or more of the feature difference clusters, wherein one of the boundaries is calculated using a 2-means event clustering algorithm to partition at least one of said feature difference clusters and at least one other boundary is calculated using a different clustering algorithm based upon a variance metric. - View Dependent Claims (21, 22, 23, 24, 25)
-
-
26. An apparatus for organizing a database of image files classified into events based upon date-time information associated with each of said image files, said apparatus comprising:
-
means for determining feature differences between consecutive said digital records; means for ranking said feature differences to provide an unclustered feature difference set; means for computing a sequence of three or more mutually exclusive feature difference clusters of said feature differences in decreasing order of relative likelihood of respective said feature differences representing separations between events, wherein at least one boundary between feature difference clusters is constrained to fall within a specified range, the feature difference clusters being computed using a 2-means event clustering algorithm to determine at least one of said feature difference clusters and using a different clustering algorithm based upon a variance metric to determine at least one other of said feature difference clusters; and means for designating respective said feature differences in the feature difference cluster lowest in said decreasing order as being within events and designating respective said feature differences in the feature difference cluster highest in said decreasing order as being between events.
-
Specification