Training an attentional cascade
First Claim
1. A system comprising:
- one or more processors configured to perform operations including;
training an attentional cascade, the attentional cascade being an ordered sequence of detector functions recorded on a machine-readable storage device, the detector functions being functions that when executed by a computer, examine a target image and return a positive result if the target image resembles an object of interest and a negative result if the target image does not resemble the object of interest, where a positive result from one detector function leads to consideration of the target image by the next detector function and a negative result from any detector function leads to rejection of the target image, wherein training the attentional cascade includes;
training each detector function in sequence starting with the first detector function, wherein training each detector function includes;
selecting a counter-example set, the counter-example set including images not resembling the object of interest, wherein selecting the counter-example set includes selecting only images that are at least a minimum difference from an example set of images, the example set including images resembling the object of interest, and where the minimum difference is specified for a distance metric for approximating a distance between particular images and images of the example set of images; and
training the detector function using only the example set and the counter-example set.
2 Assignments
0 Petitions
Accused Products
Abstract
Methods and apparatus, including computer program products, implementing techniques for training an attentional cascade. An attentional cascade is an ordered sequence of detector functions, where the detector functions are functions that examine a target image and return a positive result if the target image resembles an object of interest and a negative result if the target image does not resemble the object of interest. A positive result from one detector function leads to consideration of the target image by the next detector function, and a negative result from any detector function leads to rejection of the target image. The techniques include training each detector function in the attentional cascade in sequence starting with the first detector function. Training a detector function includes selecting a counter-example set. Selecting a counter-example set includes selecting only images that are at least a minimum difference from an example set.
-
Citations
42 Claims
-
1. A system comprising:
one or more processors configured to perform operations including; training an attentional cascade, the attentional cascade being an ordered sequence of detector functions recorded on a machine-readable storage device, the detector functions being functions that when executed by a computer, examine a target image and return a positive result if the target image resembles an object of interest and a negative result if the target image does not resemble the object of interest, where a positive result from one detector function leads to consideration of the target image by the next detector function and a negative result from any detector function leads to rejection of the target image, wherein training the attentional cascade includes; training each detector function in sequence starting with the first detector function, wherein training each detector function includes; selecting a counter-example set, the counter-example set including images not resembling the object of interest, wherein selecting the counter-example set includes selecting only images that are at least a minimum difference from an example set of images, the example set including images resembling the object of interest, and where the minimum difference is specified for a distance metric for approximating a distance between particular images and images of the example set of images; and training the detector function using only the example set and the counter-example set. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19)
-
20. A system comprising:
one or more processors configured to perform operations including; receiving an attentional cascade, the attentional cascade being an ordered sequence of detector functions recorded on a machine-readable storage device, the detector functions being functions that when executed by a computer examine a target image and return a positive result if the target image is a member of a category of objects and a negative result if the target image is not a member of the category of objects, where a positive result from one detector function leads to consideration of the target image by the next detector function and a negative result from any detector function leads to rejection of the target image; and further training the received attentional cascade to identify specific objects of interest within the category of objects to generate a trained attentional cascade, the trained attentional cascade being an ordered sequence of detector functions that when executed by a computer examine a target image and return a positive result if the target image is one of the specific objects of interest within the category of objects and a negative result if the target image is not one of the specific objects of interest within the category of objects. - View Dependent Claims (21)
-
22. A method comprising:
-
training an attentional cascade, the attentional cascade being an ordered sequence of detector functions recorded on a machine-readable storage device, the detector functions being functions that when executed by a computer examine a target image and return a positive result if the target image resembles an object of interest and a negative result if the target image does not resemble the object of interest, where a positive result from one detector function leads to consideration of the target image by the next detector function and a negative result from any detector function leads to rejection of the target image, wherein training the attentional cascade includes; training each detector function in sequence starting with the first detector function, wherein training each detector function includes; selecting a counter-example set, the counter-example set including images not resembling the object of interest, wherein selecting the counter-example set includes selecting only images that are at least a minimum difference from an example set of images, the example set including images resembling the object of interest, and where the minimum difference is specified for a distance metric for approximating a distance between particular images and images of the example set of images; and training the detector function using only the example set and the counter-example set. - View Dependent Claims (23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40)
-
-
41. A method comprising:
-
receiving an attentional cascade, the attentional cascade being an ordered sequence of detector functions recorded on a machine-readable storage device, the detector functions being functions that when executed by a computer, examine a target image and return a positive result if the target image is a member of a category of objects and a negative result if the target image is not a member of the category of objects, where a positive result from one detector function leads to consideration of the target image by the next detector function and a negative result from any detector function leads to rejection of the target image; and further training the received attentional cascade to identify specific objects of interest within the category of objects to generate a trained attentional cascade, the trained attentional cascade being an ordered sequence of detector functions that when executed by a computer examine a target image and return a positive result if the target image is one of the specific objects of interest within the category of objects and a negative result if the target image is not one of the specific objects of interest within the category of objects. - View Dependent Claims (42)
-
Specification