Object detection device, learning method, and recording medium

1. An object detection device comprising:

a memory configured to store instructions; and

one or more processors configured to execute the instructions to:

output, by a plurality of object detection units, a score indicating a probability that a predetermined object exists, for each partial region set with respect to image data inputted;

compute a weight for each of the plurality of object detection units by using weight computation parameters and based on the image data, the weight being used when the scores outputted by the plurality of object detection units are merged;

to change the weight for a predetermined object detection unit, among the computed weights, to 0 and output the weights;

merge the scores outputted by the plurality of object detection units for each of the partial regions, by using the computed weights and including the changed weight;

compute a difference between a ground truth label of the image data and the merged score as a loss; and

correct the weight computation parameters so as to reduce the loss.

2. The object detection device according to claim 1 , wherein the one or more processors are configured to normalize the weights which has not been changed to 0, among the computed weights, so that the sum of the weights becomes 1, and output the weights.

3. The object detection device according to claim 1 , wherein the one or more processors are configured to perform redistribution of the weights with a predetermined probability.

4. The object detection device according to claim 1 , wherein the one or more processors are configured to change the weight, to 0, for the object detection unit that is learned and constructed using the image data or a part of the image data inputted when the object detection device is learned.

5. The object detection device according to claim 1 ,

wherein the one or more processors are configured to compute a single weight with respect to the image data as a whole, and

wherein the one or more processors are configured to merge the scores outputted by the plurality of object detection units by using the single weight.

6. The object detection device according to claim 1 ,

wherein the one or more processors are configured to compute the weight for each partial region of the image data, and

wherein the one or more processors are configured to merge the scores outputted by the plurality of object detection units by using the weight computed for each partial region.

7. The object detection device according to claim 1 ,

wherein the one or more processors are configured to compute the weight for each class indicating the object, and

wherein the one or more processors are configured to merge the scores outputted by the plurality of object detection units by using the weight computed for each class.

8. The object detection device according to claim 1 , wherein the one or more processors are configured to multiply the scores outputted by the plurality of object detection units by the computed weight for each object detection unit, add the multiplied scores together, and calculate an average value of the multiplied scores.

9. The object detection device according to claim 1 ,

wherein the plurality of object detection units are configured to output coordinate information about a rectangular region where the object exists for each partial region,

wherein the one or more processors are configured to merge the coordinate information about the rectangular region where the object exists by using the computed weight, and

wherein the one or more processors are configured to compute a loss including a difference between the ground truth label and the coordinate information merged.

10. The object detection device according to claim 9 , wherein the one or more processors are configured to multiply the coordinate information outputted by the plurality of object detection units by the computed weight for each object detection unit, add the multiplied scores together, and calculate an average value of the multiplied scores.

11. The object detection device according to claim 1 ,

wherein the one or more processors are configured to use shooting environment prediction parameters to predict a shooting environment of the image data, and output predicted environment information,

wherein the one or more processors are configured to compute a shooting environment prediction loss on a basis of shooting environment information about the image data prepared in advance and the predicted environment information, and

wherein the one or more processors are configured to correct the shooting environment prediction parameters so as to reduce the prediction loss.

12. The object detection device according to claim 11 , wherein the one or more processors are provided with a first network including the weight computation parameters and a second network including the shooting environment prediction parameters, and

wherein the first network and the second network have a portion shared in common.

13. An object detection device comprising:

a memory configured to store instructions; and

one or more processors configured to execute the instructions to:

output, by a plurality of object detection units, a score indicating a probability that a predetermined object exists, for each partial region set with respect to image data inputted;

compute weights used for merging the scores outputted by the plurality of object detection units by using weight computation parameters and based on the image data;

change the weight for a predetermined object detection unit, among the computed weights, to 0 and output the weights;

merge the scores outputted by the plurality of object detection units for each of the partial regions, by using the computed weights and including the changed weight;

output, by an object detection unit of a target model, a score indicating the probability that the predetermined object exists, for each partial region set with respect to the image data;

compute a loss indicating a difference of the output of the object detection unit of the target model from a ground truth label of the image data and the merged score; and

correct the parameters of the object detection unit of the target model so as to reduce the loss.

14. A learning method of an object detection device, comprising:

outputting, by a plurality of object detection units, a score indicating a probability that a predetermined object exists, for each partial region set with respect to image data inputted;

computing a weight for each of the plurality of object detection units by using weight computation parameters and based on the image data, the weight being used when the scores outputted by the plurality of object detection units are merged;

changing the weight for a predetermined object detection unit, among the computed weights, to 0 and outputting the weights;

merging the scores outputted by the plurality of object detection units for each of the partial regions, by using the computed weights including the changed weight;

computing a difference between a ground truth label of the image data and the merged score as a loss; and

correcting the weight computation parameters so as to reduce the loss.

15. A non-transitory computer-readable recording medium having recorded thereon instructions executable by a processor to perform the method of claim 14 .

16. A learning method of an object detection device, comprising:

outputting, by a plurality of object detection units, a score indicating a probability that a predetermined object exists, for each partial region set with respect to image data inputted;

computing weights used for merging the scores outputted by the plurality of object detection units by using weight computation parameters and based on the image data;

changing the weight for a predetermined object detection unit, among the computed weights, to 0 and outputting the weights;

merging the scores outputted by the plurality of object detection units for each of the partial regions, by using the computed weights including the changed weight;

outputting, by an object detection unit of a target model, a score indicating the probability that the predetermined object exists, for each partial region set with respect to the image data;

computing a loss indicating a difference of the output of the object detection unit of the target model from a ground truth label of the image data and the merged score; and

correcting the parameters of the object detection unit of the target model so as to reduce the loss.

17. A non-transitory computer-readable recording medium having recorded thereon instructions executable by a processor to perform the method of claim 16 .