Granted Patent B2

US 12,738,007 · App. 18/399,031 · Granted Sep 15, 2026

Inventor: Soma Shiraishi (Tokyo, JP)

Assignee: NEC CORPORATION

G06V10/25G06V10/764G06V20/70G06V2201/07

Abstract

Claims (50)

1 . An information processing apparatus comprising

at least one processor, the at least one processor carrying out:

an acquiring process of acquiring a training image in which a region containing an image of an object has a ground-truth label attached thereto;

a score assigning process of assigning, to each of unit regions, a score which indicates objectness, by inputting the training image to a region detection model for detecting, in an image, a region corresponding to an image of an object;

a first pseudo label assigning process of assigning a pseudo label which indicates an object, to a partial region of the region having the ground-truth label attached thereto, in accordance with the score assigned to each of the unit regions;

a second pseudo label assigning process of assigning a pseudo label which indicates a background, to a partial region of a region not having the ground-truth label attached thereto, in accordance with the score assigned to each of the unit regions;

an updating process of updating a model parameter of the region detection model, with reference to a region having attached thereto the pseudo label which indicates an object or the pseudo label which indicates a background;

a detecting process of inputting a captured image to the region detection model with the updated model parameter to detect an object region contained in the captured image, and outputting information representing the detected object region for display on a terminal,

wherein in the first pseudo label assigning process, the at least one processor assigns the pseudo label which indicates an object, to a unit region having been assigned the score that falls within a range of scores including a highest score, the range corresponding to a predetermined proportion, in the region having the ground-truth label attached thereto,

wherein the training image contains a plurality of regions each being the region containing an image of an object, and

wherein in the first pseudo label assigning process, the at least one processor assigns the pseudo label which indicates an object, to the unit region having been assigned the score that falls within a range of scores including a highest score, the range corresponding to a predetermined proportion, in each of the plurality of regions each of which contains an image of an object.

2 . The information processing apparatus according to claim 1 , wherein

in the second pseudo label assigning process, the at least one processor assigns the pseudo label which indicates a background, to a unit region having been assigned the score that falls within a range of scores including a lowest score, the range corresponding to a predetermined proportion, in the region not having the ground-truth label attached thereto.

3 . The information processing apparatus according to claim 1 , wherein

the ground-truth label is a class label,

in the score assigning process, the at least one processor assigns a score for each of classes, to each of the unit regions contained in the training image, and

in the first pseudo label assigning process, the at least one processor assigns a pseudo label which indicates one of the classes, to the partial region of the region having the ground-truth label attached thereto, in accordance with the score assigned to each of the unit regions.

4 . The information processing apparatus according to claim 1 , wherein

in the first pseudo label assigning process, the at least one processor:

assigns the pseudo label which indicates an object, to the partial region of the region having the ground-truth label attached thereto, in accordance with the score assigned to each of the unit regions; and

further assigns the pseudo label which indicates an object, to the partial region of the region not having the ground-truth label attached thereto, in accordance with the score assigned to each of the unit regions.

5 . The information processing apparatus according to claim 4 , wherein

the at least one processor further carries out:

an identification label assigning process of assigning, to each of the unit regions contained in the region having the ground-truth label attached thereto, an identification label which indicates an object or a background, by inputting at least a part of the training image to a second region detection model for detecting, in an image, a region corresponding to an image of an object; and

a threshold determining process of determining a threshold in accordance with a score distribution of the unit regions each having attached thereto the identification label which indicates an object and a score distribution of the unit regions each having attached thereto the identification label which indicates a background, and

in the first pseudo label assigning process, the at least one processor assigns the pseudo label which indicates an object, to a unit region having been assigned the score which is equal to or greater than the threshold in the training image, and

in the second pseudo label assigning process, the at least one processor assigns the pseudo label which indicates a background, to a unit region having been assigned the score which is smaller than the threshold or a second threshold smaller than the threshold in the training image.

6 . The information processing apparatus according to claim 5 , wherein

in the threshold determining process, the at least one processor uses, as the threshold, a value obtained by adding a predetermined value to an average of a center of gravity of the score distribution of the unit regions each having attached thereto the identification label which indicates an object and a center of gravity of the score distribution of the unit regions each having attached thereto the identification label which indicates a background, and uses, as the second threshold, a value obtained by subtracting the predetermined value from the average, and

in the second pseudo label assigning process, the at least one processor assigns the pseudo label which indicates a background, to a unit region having been assigned the score which is smaller than the second threshold in the training image.

7 . An information processing method comprising:

at least one processor acquiring a training image in which a region containing an image of an object has a ground-truth label attached thereto;

the at least one processor assigning, to each of unit regions, a score which indicates objectness, by inputting the training image to a region detection model for detecting, in an image, a region corresponding to an image of an object;

the at least one processor assigning a pseudo label which indicates an object, to a partial region of the region having the ground-truth label attached thereto, in accordance with the score assigned to each of the unit regions;

the at least one processor assigning a pseudo label which indicates a background, to a partial region of a region not having the ground-truth label attached thereto, in accordance with the score assigned to each of the unit regions;

the at least one processor updating a model parameter of the region detection model, with reference to a region having attached thereto the pseudo label which indicates an object or the pseudo label which indicates a background; and

the at least one processor inputting a captured image to the region detection model with the updated model parameter to detect an object region contained in the captured image, and outputting information representing the detected object region for display on a terminal,

wherein in assigning the pseudo label which indicates an object, the at least one processor assigns the pseudo label which indicates an object, to a unit region having been assigned the score that falls within a range of scores including a highest score, the range corresponding to a predetermined proportion, in the region having the ground-truth label attached thereto,

wherein the training image contains a plurality of regions each being the region containing an image of an object, and

wherein in assigning the pseudo label which indicates an object, the at least one processor assigns the pseudo label which indicates an object, to the unit region having been assigned the score that falls within a range of scores including a highest score, the range corresponding to a predetermined proportion, in each of the plurality of regions each of which contains an image of an object.

8 . A computer-readable, non-transitory recording medium having recorded thereon a program for causing a computer to carry out:

a process of acquiring a training image in which a region containing an image of an object has a ground-truth label attached thereto;

a process of assigning, to each of unit regions, a score which indicates objectness, by inputting the training image to a region detection model for detecting, in an image, a region corresponding to an image of an object;

a process of assigning a pseudo label which indicates an object, to a partial region of the region having the ground-truth label attached thereto, in accordance with the score assigned to each of the unit regions;

a process of assigning a pseudo label which indicates a background, to a partial region of a region not having the ground-truth label attached thereto, in accordance with the score assigned to each of the unit regions;

a process of updating a model parameter of the region detection model, with reference to a region having attached thereto the pseudo label which indicates an object or the pseudo label which indicates a background; and

a process of inputting a captured image to the region detection model with the updated model parameter to detect an object region contained in the captured image, and outputting information representing the detected object region for display on a terminal,

wherein in the process of assigning a pseudo label which indicates an object, the computer assigns the pseudo label which indicates an object, to a unit region having been assigned the score that falls within a range of scores including a highest score, the range corresponding to a predetermined proportion, in the region having the ground-truth label attached thereto,

wherein the training image contains a plurality of regions each being the region containing an image of an object, and

wherein in the process of assigning a pseudo label which indicates an object, the computer assigns the pseudo label which indicates an object, to the unit region having been assigned the score that falls within a range of scores including a highest score, the range corresponding to a predetermined proportion, in each of the plurality of regions each of which contains an image of an object.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (16)