| CPC G06T 7/73 (2017.01) [G06N 20/00 (2019.01); G06T 7/66 (2017.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/20164 (2013.01)] | 7 Claims |
|
1. A learning device comprising:
a processor; and
a memory storing instructions executable by the processor to:
train, on a basis of a first training image of an object and a training label including a correct coordinate value relating to a feature point of the object, a first discriminator configured to output, from a first input image, a predicted coordinate value relating to a feature point in the first input image;
input the first training image to the trained first discriminator and receive as output from the first discriminator a predicted coordinate value relating to the feature point in the first training image;
determine, on a basis of the predicted coordinate value output by the first discriminator from the first training image, a candidate area of the feature point in the first training image;
generate a second training image by cropping the first training image to the candidate area as centered on the predicted coordinate value; and
train, on a basis of the second training image, a second discriminator configured to output, from a second input image, a reliability map indicating a reliability of a feature point at each of a plurality of blocks in the second input image,
wherein the processor selects, from multiple candidates of a plurality of the correct coordinate values, a candidate for which an accuracy of the reliability map for the candidate area is highest, and trains, on a basis of the selected candidate and the training image, the first discriminator,
wherein the processor trains, on a basis of the training image and the plurality of the correct coordinate values, the first discriminator configured to output a plurality of the predicted coordinate values, and
wherein the processor determines the candidate area from the plurality of the predicted coordinate values.
|
|
3. A discrimination device comprising:
a processor; and
a memory storing instructions executable by the processor to:
input a first training image of an object to a first discriminator and receive as output from the first discriminator a predicted coordinate value relating to a feature point of the object in the first training image, the first discriminator trained on a basis of the first training image and a training label including a correct coordinate valuing relating to the feature point, the first discriminator configured to output, from a first input image, a predicted coordinate value relating to a feature point in the first input image;
determine, on a basis of the predicted coordinate value output by the first discriminator from the first training image, a candidate area of the feature point in the first training image;
generate a second training image by cropping the first training image to the candidate area as centered on the predicted coordinate value; and
input the second training image to a second discriminator and receive as output from the second discriminator a reliability map indicating a reliability of the feature point at each of a plurality of blocks in the second training image, the second discriminator trained and configured to output, from a second input image, a reliability map indicating a reliability of a feature point at each of a plurality of blocks in the second input image,
wherein the processor determines the candidate area based on an area parameter read from a storage device that stores the area parameter and the predicted coordinate value output by the first discriminator, the area parameter used to determine the candidate area from the predicted coordinate value.
|