SYSTEM AND METHOD OF CHAINING ALGORITHMS FOR GLOBAL OBJECT RECOGNITION TO IMPROVE PROBABILITY OF CORRECTNESS AND REDUCE PROCESSING LOAD
First Claim
1. A computer-implemented method for global object recognition comprising:
- receiving, by the one or more hardware processors, object metadata including a plurality of characteristics that define an object to be detected;
receiving, by one or more hardware processors, search metadata including a plurality of context parameters that define a search for the object;
retrieving, based on the object and search metadata, a plurality of source data of a given data type;
selecting, from a plurality of algorithms, a subset of algorithms to be used in processing the retrieved source data based on a cumulative trained probability of correctness (Pc) that each of the algorithms, which are processed in a chain and conditioned upon the result of the preceding algorithms, produce a correct result;
ordering the algorithms in the subset based on algorithm metadata including a plurality of algorithm characteristics to reduce an expected processing load of the retrieved source data; and
processing the retrieved source data in order according to the chain of the selected subset of algorithms to obtain a plurality of results and to reduce the number of source data that is processed by the next algorithm in the chain, at least one result indicating whether the object was detected in corresponding source data output from the last algorithm in the chain.
1 Assignment
0 Petitions
Accused Products
Abstract
A system and method improves the probability of correctly detecting an object from a collection of source data and reduces the processing load. A plurality of algorithms for a given data type are selected and ordered based on a cumulative trained probability of correctness (Pc) that each of the algorithms, which are processed in a chain and conditioned upon the result of the preceding algorithms, produce a correct result and a processing. The algorithms cull the source data to pass forward a reduced subset of source data in which the conditional probability of detecting the object is higher than the a priori probability of the algorithm detecting that same object. The Pc and its confidence interval is suitably computed and displayed for each algorithm and the chain and the final object detection.
-
Citations
22 Claims
-
1. A computer-implemented method for global object recognition comprising:
-
receiving, by the one or more hardware processors, object metadata including a plurality of characteristics that define an object to be detected; receiving, by one or more hardware processors, search metadata including a plurality of context parameters that define a search for the object; retrieving, based on the object and search metadata, a plurality of source data of a given data type; selecting, from a plurality of algorithms, a subset of algorithms to be used in processing the retrieved source data based on a cumulative trained probability of correctness (Pc) that each of the algorithms, which are processed in a chain and conditioned upon the result of the preceding algorithms, produce a correct result; ordering the algorithms in the subset based on algorithm metadata including a plurality of algorithm characteristics to reduce an expected processing load of the retrieved source data; and processing the retrieved source data in order according to the chain of the selected subset of algorithms to obtain a plurality of results and to reduce the number of source data that is processed by the next algorithm in the chain, at least one result indicating whether the object was detected in corresponding source data output from the last algorithm in the chain. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20)
-
-
21. A computer-implemented method for global object recognition comprising:
-
receiving, by the one or more hardware processors, object metadata including a plurality of characteristics that define the object to be located; receiving, by one or more hardware processors, for each of a plurality of algorithms configured to process source data of a given data type, algorithm metadata including a plurality of algorithm characteristics that describe the algorithm; retrieving a plurality of training source data of the given data type; selecting, from the plurality of algorithms, based on the object and algorithm metadata a plurality of candidate subsets of algorithms to be used in processing the retrieved source data; for each candidate subset, ordering the algorithms in the chain based on algorithm metadata to reduce an expected processing load; for each candidate subset, processing the retrieved source data in order according to the chain of algorithms to obtain a plurality of results and to reduce the number of training source data that is processed by the next algorithm in the chain, at least one result indicating whether the object was identified in corresponding source data output from the last algorithm in the chain; for each candidate subset, computing a cumulative trained probability of correctness (Pc) and corresponding confidence interval that each of the algorithms, which are processed in the chain and conditioned upon the result of the preceding algorithms, produce a correct result; selecting a candidate subset based on its trained Pc and corresponding confidence interval and expected processing load; pairing the selected candidate subset of algorithms with the object to be detected; and repeating the steps for a plurality of different objects.
-
-
22. A computer-implemented method for global object recognition comprising:
-
receiving, by the one or more hardware processors, object metadata including a plurality of characteristics that define the object to be detected; receiving, by one or more hardware processors, search metadata including a plurality of context parameters that define a search for the object; receiving, by one or more hardware processors, a plurality algorithms, algorithm metadata including a plurality of algorithm characteristics that describe each algorithm, and a plurality of defined subsets of chained algorithms configured to detect different objects, each said defined subset selected based on a cumulative trained probability of correctness (Pc) and corresponding confidence interval that each one of the algorithms, which are processed in the chain and conditioned upon the result of the preceding algorithms, produce a correct result and an expected processing load of the chain; selecting, from the plurality of defined subsets, based on the object metadata one of the defined subsets, said algorithms in the selected subsets configured to process source data of a given data type; if none of the defined subsets match the object to be located, based on the object and algorithm metadata selecting and ordering a plurality of algorithms, configured to process source data of a given data type, to define a selected subset; retrieving, based on the plurality of context parameters, a plurality of source data of the given data type; processing the retrieved one or more source data in order according to the chain of the selected subset of algorithms to obtain a plurality of results and to reduce the number of source data that is processed by the next algorithm in the chain, at least one result indicating whether the object was detected in corresponding source data output from the last algorithm in the chain; and determining a cumulative object Pc and confidence interval representing whether the object was detected in one or more of the retrieved source data output from the last algorithm based on the at least one result.
-
Specification