Label estimation device, label estimation method, and label estimation program

Patent metadata

Field	Value
Publication number	US-12136252-B2
Application number	US-201917628071-A
Country	US
Kind code	B2
Filing date	Jul 19, 2019
Priority date	Jul 19, 2019
Publication date	Nov 5, 2024
Grant date	Nov 5, 2024

Abstract

A point group including a small number of points that have been assigned labels is taken as an input to assign labels to points that have not been assigned labels.
In a label estimation apparatus for estimating a label to be assigned to a point that has not been labeled using a label of a point that has been labeled among points included in a point group, a confidence derivation unit 103 takes a point that has not been labeled within a point group including a point that has been labeled and the point that has not been labeled as a target point and estimates a class of the target point and a likelihood indicating a confidence of an estimation result of the class from a set of points included in the point group, a priority derivation unit 104 obtains a distance between the target point and a point that has been assigned the same label as a label corresponding to the estimated class as a priority used to determine whether the estimated class is appropriate, and a label determination unit 105 determines whether the estimated class is appropriate using at least an index based on the distance.

First claim

The invention claimed is:

1. A label estimation apparatus configured to estimate a label to be assigned to a point that has not been labeled using a label of a point that has been labeled among points included in a point group, the label estimation apparatus comprising circuit configured to execute a method comprising:
taking, as a target point, a point that has not been labeled within a point group including a point that has been labeled and the point that has not been labeled;
estimating a class of the target point and a likelihood indicating a confidence of an estimation result of the class from a set of points included in the point group;
obtaining a distance between the target point and a point that has been assigned the same label as a label corresponding to the estimated class as a priority used to determine whether the estimated class is appropriate; and
determining whether the estimated class is appropriate using at least an index based on the distance,
wherein the method further comprising:
learning a model that derives a likelihood that an input point is classified as each of a plurality of classes while associating a coordinate of each of a plurality of points with a label indicating which class each of the plurality of points is classified as, and
learning the model to minimize a loss function, the loss function including a first term regarding an error for the point that has been labeled in advance and a second term regarding an error for the point that has been assigned the label indicating the class that has been determined to be appropriate and weighting the first and second terms such that contributions of the point that has been labeled in advance and the point that has been assigned the label indicating the class that has been determined to be equal.

2. The label estimation apparatus according to claim 1 , wherein a set of points that is input to the obtaining includes only points present within a predetermined distance from the target point.

3. The label estimation apparatus according to claim 1 , wherein the index based on the distance is designed such that it becomes easy to determine that the estimated class is appropriate as the distance decreases.

4. The label estimation apparatus according to claim 1 , the circuit further configured to execute a method comprising:
determining whether the class is appropriate based on the index based on the distance and the likelihood.

5. The label estimation apparatus according to claim 1 , the circuit further configured to execute a method comprising:
repeating learning of the model using coordinates of a point that has been labeled in advance and a point that has been assigned a label indicating a class that has been determined to be appropriate by the label determination unit; and
repeating estimation of the class and the likelihood of the target point using the model learned by iterative processing.

6. A label estimation method for estimating a label to be assigned to a point that has not been labeled using a label of a point that has been labeled among points included in a point group, the label estimation method comprising:
taking a point that has not been labeled within a point group including a point that has been labeled and the point that has not been labeled as a target point and estimating a class of the target point and a likelihood indicating a confidence of an estimation result of the class from a set of points included in the point group;
obtaining a distance between the target point and a point that has been assigned the same label as a label corresponding to the estimated class as a priority used to determine whether the estimated class is appropriate; and
determining whether the estimated class is appropriate using at least an index based on the distance,
wherein the method further comprising:
learning a model that derives a likelihood that an input point is classified as each of a plurality of classes while associating a coordinate of each of a plurality of points with a label indicating which class each of the plurality of points is classified as, and
learning the model to minimize a loss function, the loss function including a first term regarding an error for the point that has been labeled in advance and a second term regarding an error for the point that has been assigned the label indicating the class that has been determined to be appropriate and weighting the first and second terms such that contributions of the point that has been labeled in advance and the point that has been assigned the label indicating the class that has been determined to be equal.

7. A computer-readable non-transitory recording medium storing computer-executable label estimation program instructions that when executed by a processor cause a computer system to execute a method to estimate a label to be assigned to a point that has not been labeled using a label of a point that has been labeled among points included in a point group, comprising:
taking, as a target point, a point that has not been labeled within a point group including a point that has been labeled and the point that has not been labeled and estimate a class of the target point and a likelihood indicating a confidence of an estimation result of the class from a set of points included in the point group;
obtaining a distance between the target point and a point that has been assigned the same label as a label corresponding to the estimated class as a priority used to determine whether the estimated class is appropriate; and
determining whether the estimated class is appropriate using at least an index based on the distance,
wherein the method further comprising:
learning a model that derives a likelihood that an input point is classified as each of a plurality of classes while associating a coordinate of each of a plurality of points with a label indicating which class each of the plurality of points is classified as, and
learning the model to minimize a loss function, the loss function including a first term regarding an error for the point that has been labeled in advance and a second term regarding an error for the point that has been assigned the label indicating the class that has been determined to be appropriate and weighting the first and second terms such that contributions of the point that has been labeled in advance and the point that has been assigned the label indicating the class that has been determined to be equal.

8. The label estimation apparatus according to claim 2 , wherein the index based on the distance is designed such that it becomes easy to determine that the estimated class is appropriate as the distance decreases.

9. The label estimation method according to claim 6 , wherein a set of points that is input to the obtaining includes only points present within a predetermined distance from the target point.

10. The label estimation method according to claim 6 , wherein the index based on the distance is designed such that it becomes easy to determine that the estimated class is appropriate as the distance decreases.

11. The label estimation method according to claim 6 , the method further comprising:
determining whether the class is appropriate based on the index based on the distance and the likelihood.

12. The label estimation method according to claim 7 , the method further comprising:
repeating learning of the model using coordinates of a point that has been labeled in advance and a point that has been assigned a label indicating a class that has been determined to be appropriate by the label determination unit; and
repeating estimation of the class and the likelihood of the target point using the model

Assignees

• Nippon Telegraph & Telephone

Inventors

• Yao Yasuhiro
• Murasaki Kazuhiko
• Ando Shingo
• Sagata Atsushi

Classifications

• G06V10/776 Primary

  Validation; Performance evaluation · CPC title

• G06V10/7747

  Organisation of the process, e.g. bagging or boosting · CPC title

• G06V10/764 Primary

  using classification, e.g. of video objects · CPC title

• G06V20/00

  Scenes; Scene-specific elements (control of digital cameras H04N23/60) · CPC title

• G06V10/761

  Proximity, similarity or dissimilarity measures · CPC title