Image processing apparatus, training apparatus, image processing method, training method, and non-transitory computer-readable medium
AssigneeCanon Kabushiki Kaisha
InventorYuta Horikawa
Life of this patent
Estimating the term…
Abstract
An image processing apparatus is provided. The image processing apparatus searches for a tracking target from a target image. A feature of the target image and a feature of a first image of the tracking target at a first time are acquired. A weight of a feature for each of a plurality of positions of the first image of the tracking target is generated on the basis of the first image of the tracking target and a second image of the tracking target at a second time. The tracking target is detected from the target image on the basis of a correlation between a feature of the target image and a feature of a first image of the tracking target weighted on the basis of the weight.
Claims22
An image processing apparatus that searches for a tracking target from a target image, the image processing apparatus comprising one or more memories storing instructions and one or more processors that execute the instructions to:
- acquire a feature map of the target image and a feature map of a first image of the tracking target at a first time;
- generate a weight map of the feature map of the first image of the tracking target on the basis of the first image of the tracking target and a second image of the tracking target at a second time, wherein the weight map reflects a magnitude of a time-series change of the image of the tracking target for each position of the first image;
- and
- detect the tracking target from the target image on the basis of a correlation between the feature map of the target image and the feature map of the first image of the tracking target weighted on the basis of the weight map.
An image processing apparatus that searches for a tracking target from a target image, the image processing apparatus comprising one or more memories storing instructions and one or more processors that execute the instructions to:
- acquire a feature map of the target image and a feature map of a first image of the tracking target at a first time;
- generate a weight map of the feature map of the first image of the tracking target on the basis of a time-series change in an image of the tracking target;
- and
- detect the tracking target from the target image on the basis of a correlation between the feature map of the target image and the feature map of the first image of the tracking target weighted on the basis of the weight map.
A training apparatus that trains a weight map generation model, the training apparatus comprising one or more memories storing instructions and one or more processors that execute the instructions to:
- acquire a weight map of a feature map of a first training image obtained by inputting the first training image of a tracking target, a second training image of the tracking target, and the feature map of the first training image to the weight map generation model;
- evaluate an error between the weight map of the feature map of the first training image and supervisory data of the weight map of the feature map of the first training image;
- and
- update the weight map generation model based on an evaluation result of the error,
- wherein the weight map generation model is used by an image processing apparatus that searches for a tracking target from a target image, the image processing apparatus comprising one or more memories storing instructions and one or more processors that execute the instructions to:
- acquire a feature map of the target image and a feature map of a first image of the tracking target at a first time;
- generate a weight map indicating a weight of the feature map for each of a plurality of positions of the first image of the tracking target on the basis of the first image of the tracking target and a second image of the tracking target at a second time using the weight map generation model;
- and
- detect the tracking target from the target image on the basis of a correlation between the feature map of the target image and the feature map of the first image of the tracking target weighted on the basis of the weight map.
An image processing method for searching for a tracking target from a target image, comprising:
- acquiring a feature map of the target image and a feature map of a first image of the tracking target at a first time;
- generating a weight map of the feature map of the first image of the tracking target on the basis of the first image of the tracking target and a second image of the tracking target at a second time, wherein the weight map reflects a magnitude of a time-series change of the image of the tracking target for each position of the first image;
- and
- detecting the tracking target from the target image on the basis of a correlation between the feature map of the target image and the feature map of the first image of the tracking target weighted on the basis of the weight map.
A non-transitory computer-readable medium storing a program executable by a computer to perform a method for searching for a tracking target from a target image, the method comprising:
- acquiring a feature map of the target image and a feature map of a first image of the tracking target at a first time;
- generating a weight map of the feature map of the first image of the tracking target on the basis of the first image of the tracking target and a second image of the tracking target at a second time, wherein the weight map reflects a magnitude of a time-series change of the image of the tracking target for each position of the first image;
- and
- detecting the tracking target from the target image on the basis of a correlation between the feature map of the target image and the feature map of the first image of the tracking target weighted on the basis of the weight map.
A training method for training a weight map generation model, comprising:
- acquiring a weight map of a feature map of a first training image obtained by inputting the first training image of a tracking target, a second training image of the tracking target, and the feature map of the first training image to the weight map generation model;
- evaluating an error between the weight map of the feature map of the first training image and supervisory data of the weight map of the feature map of the first training image;
- and
- updating the weight map generation model based on an evaluation result of the error,
- wherein the weight map generation model is used in an image processing method for searching for a tracking target from a target image, the image processing method comprising:
- acquiring a feature map of the target image and a feature map of a first image of the tracking target at a first time;
- generating a weight map indicating a weight of the feature map for each of a plurality of positions of the first image of the tracking target on the basis of the first image of the tracking target and a second image of the tracking target at a second time using the weight map generation model;
- and
- detecting the tracking target from the target image on the basis of a correlation between the feature map of the target image and the feature map of the first image of the tracking target weighted on the basis of the weight map.
A non-transitory computer-readable medium storing a program executable by a computer to perform a method for training a weight map generation model, the method comprising:
- acquiring a weight map of a feature map of a first training image obtained by inputting the first training image of a tracking target, a second training image of the tracking target, and the feature map of the first training image to the weight map generation model;
- evaluating an error between the weight map of the feature map of the first training image and supervisory data of the weight map of the feature map of the first training image;
- and
- updating the weight map generation model based on an evaluation result of the error,
- wherein the weight map generation model is used in an image processing method for searching for a tracking target from a target image, the image processing method comprising:
- acquiring a feature map of the target image and a feature map of a first image of the tracking target at a first time;
- generating a weight map indicating a weight of the feature map for each of a plurality of positions of the first image of the tracking target on the basis of the first image of the tracking target and a second image of the tracking target at a second time using the weight map generation model;
- and
- detecting the tracking target from the target image on the basis of a correlation between the feature map of the target image and the feature map of the first image of the tracking target weighted on the basis of the weight map.
Dependencies are inferred from the claim text.