Back to results
US 9,070,045 B2Granted
Crosstalk cascades for use in object detection
AssigneeMicrosoft Technology Licensing, LLC
InventorsPiotr Dollar, Wolf Kienzle
Life of this patent
Estimating the term…
Abstract
Technologies pertaining to object detection are described herein. A cascaded classifier executes over subwindows of an image in a plurality of stages. A crosstalk cascade is employed to reject subwindows as being candidates for including an object that is desirably detected, where the crosstalk cascade is a combination of multiple cascades.
Claims20
A method for detecting a specified object in an image, the method comprising:
- a) receiving the image;
- b) selecting a classifier from a cascade of classifiers, wherein the classifier is configured to analyze subwindows of the image and output scores that are indicative of likelihood of the object existing in respective subwindows;
- c) selecting, from candidate subwindows, a first set of subwindows, wherein a number of subwindows in the first set of subwindows is less than a number of subwindows in the candidate subwindows;
- d) for each subwindow in the first set of subwindows, executing the classifier over a respective subwindow such that the classifier outputs a respective score for the respective subwindow;
- e) for each subwindow in the first set of subwindows, comparing the respective score assigned to the respective subwindow with an excitatory cascade threshold value;
- f) for each subwindow in the first set of subwindows, when the score assigned to the respective subwindow in the first set of subwindows is below the excitatory cascade threshold value, removing the respective subwindow and any other subwindow in a defined neighborhood of the respective subwindow from the candidate set;
- g) subsequent to the removing of the respective subwindow and any other subwindows in the defined neighborhood of the respective subwindow, executing the classifier over a second set of subwindows, the second set of subwindows comprising subwindows in the candidate set except for any subwindows in the first set;
- h) for each subwindow in the candidate set, comparing the score assigned to the respective subwindow with a soft cascade threshold value;
- i) when the score assigned to the respective subwindow is below the soft cascade threshold value, removing the respective subwindow from the candidate set;
- and
- j) detecting the specified object in the image as being in at least one subwindow from the candidate set.
8 more elements
A method that facilitates detecting an object in an image, the method comprising:
- executing a cascaded classifier over subwindows of the image, the cascaded classifier comprising a sequence of weak classifiers that correspond to a respective sequence of stages, wherein accuracy of the cascaded classifier increases as the sequence of stages advances, and wherein at each stage in the sequence of stages a respective weak classifier outputs scores for respective subwindows of the image, wherein a score assigned to a subwindow is indicative of a probability that the respective subwindow includes the object;
- in multiple stages, rejecting at least one subwindow from the subwindows based upon a score assigned to a neighbor of the subwindow in the image by the cascaded classifier;
- and
- detecting the object in the image based at least in part upon the rejecting of the at least one subwindow.
1 more element
A computing device that is programmed to detect an object in an image, the computing device comprising:
- a processor;
- and
- a memory that comprises instructions that are executable by the processor, the instructions comprising:
- first instructions for forming a candidate set of subwindows in the image by rejecting at least a first subwindow based upon a score assigned to a second subwindow;
- second instructions for removing a third subwindow from the candidate set based upon a score assigned to a fourth subwindow in the candidate set, the score assigned to the second subwindow and the score assigned to the fourth subwindow are generated by a cascaded classifier;
- and
- third instructions for indicating that the object is in the image based upon the score assigned to the fourth subwindow.
4 more elements
Dependencies are inferred from the claim text.