Identifying descriptor for person and object in an image
View Patent ↗A device and method for processing an image to create appearance and shape labeled images of a person or object captured within the image. The appearance and shape labeled images are unique properties of the person or object and can be used to re-identify the person or object in subsequent images. The appearance labeled image is an aggregate of pre-stored appearance labels that are assigned to image segments of the image based on calculated appearance attributes of each image segment. The shape labeled image is an aggregate of pre-stored shape labels that are assigned to image segments of the image based on calculated shape attributes of each image segment. An identifying descriptor of the person or object can be computed based on both the appearance labeled image and the shape labeled image. The descriptor can be compared with other descriptors of later captured images to re-identify a person or object.
1. A method for processing an image that includes a person or object that had been identified in the image, the method including:
identifying a plurality of image segments that divide the person or object in the image into portions that capture the person or object;
calculating an appearance attribute for each of the image segments;
creating an appearance labeled image in an appearance labeling device by assigning appearance labels to the image segments based on the calculated appearance attribute of each image segment, where the appearance labels correspond to representative appearance attributes predetermined from an appearance image database; and
generating an identifying descriptor for the person or object that captures a spatial relationship among the appearance labels in the appearance labeled image.
2. A method for processing an image that includes a person or object that had been identified in the image, the method including:
identifying a plurality of image segments that divide the person or object in the image into portions that capture the person or object;
calculating an appearance attribute for each of the image segments;
creating an appearance labeled image in an appearance labeling device by assigning appearance labels to the image segments based on the calculated appearance attribute of each image segment, where the appearance labels correspond to representative appearance attributes predetermined from an appearance image database;
calculating a histogram of oriented gradients (HOG) for each of the plurality of image segments; and
including the HOG in the appearance attribute.
3. The method of claim 2 , wherein the HOG is calculated for an N×N patch of pixels within an image segment and N is between about 2 to 25.
4. The method of claim 3 , wherein a gradient within each patch is quantized into about 4 to 28 directions.
5. The method of claim 3 , wherein a gradient within each patch is quantized into about 8 directions.
6. A method for processing an image that includes a person or object that had been identified in the image, the method including:
identifying a plurality of image segments that are included as part of the person or object in the image;
calculating an appearance attribute for each of the image segments;
creating an appearance labeled image in an appearance labeling device by assigning appearance labels to the image segments based on the calculated appearance attribute of each image segment, where the appearance labels correspond to representative appearance attributes predetermined from an appearance image database;
placing a mask having a plurality of discrete regions over the plurality of image segments of the person or object;
generating a histogram of appearance labels occurring in each discrete region of the mask;
generating an average histogram by averaging histograms of appearance labels in each discrete region of the mask; and
concatenating average histograms of discrete regions of the mask to form an identifying descriptor.
7. The method of claim 6 , wherein each of the discrete regions of the mask has an orientation-dependent shape.
8. The method of claim 6 , wherein each of the discrete regions of the mask has an L-shape.