Face Track Recognition with Multi-sample Multi-view Weighting

• US 20150269421A1
• Filed: 03/19/2015
• Published: 09/24/2015
• Est. Priority Date: 03/19/2014
• Status: Active Grant

• Current Assignee: Hulu LLC (The Walt Disney Company)
• Original Assignee: Hulu LLC (The Walt Disney Company)
• Inventors: Liu, Cailiang, Wang, Zhibing, Zhang, Chenguang, Xiong, Tao
• Granted Patent: US 9,852,364 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 18/253 of extracted features
• CPC Class Codes: G06V 10/806 of extracted features
• CPC Class Codes: G06V 40/173 face re-identification, e.g. recognising unknown faces across different face tracks

Abstract

In one embodiment, a method determines known features for existing face tracks that have identity labels and builds a database using these features. The face tracks may have multiple different views of a face. Multiple features from the multiple faces may be taken to build the face models. For an unlabeled face track without identity information, the method determines its sampled features and finds labeled nearest neighbor features with respect to multiple feature spaces from the face models. For each face in the unlabeled face track, the method decomposes the face as a linear combination of its neighbors from the known features from the face models. Then, the method determines weights for the known features to weight the coefficients of the known features. Particular embodiments use a non-linear weighting function to learn the weights that provides more accurate labels.

Claims

1. A method comprising:

determining, by a computing device, a plurality of face tracks for detected faces in a video;

receiving, by the computing device, a set of labels for a portion of the plurality of face tracks from a set of users to form a set of labeled face tracks, wherein a label in the set of labels identifies an identity for a face track in the set of labeled face tracks;

extracting, by the computing device, a first set of features for an unlabeled face track in the plurality of face tracks, wherein an identity for the unlabeled face track is not known;

correlating, by the computing device, the extracted first set of features for the unlabeled face track to a second set of features from the set of labeled face tracks;

generating, by the computing device, feature weights for the first set of features based on confidence scores for the second set of features using a weighting function that magnifies feature weights for the second set of features with higher confidence scores and suppresses feature weights for the second set of features with lower confidence scores in a non-linear manner; and

using, by the computing device, the generated feature weights to determine a label for the unlabeled face track by applying the generated feature weights to the second set of features.

2. The method of claim 1, wherein generating the feature weights comprises:

selecting the second set of features for the set of labeled face tracks;

generating a feature vector for each feature in the second set of features based on the correlating; and

generating feature weights based on the feature vector for each feature in the second set of features.

3. The method of claim 2, wherein the feature vector comprises a non-linear portion and a linear portion.

4. The method of claim 1, wherein using the generated feature weights to generate the label comprises:

combining the feature weights for the first set of features to generate combined fusion scores for each face track in the portion of the plurality of face tracks; and

determining the label based on the combined fusion scores for each face track.

5. The method of claim 4, wherein a combined fusion score is generated non-linearly.

6. The method of claim 4, wherein the combined fusion score magnifies higher confidence scores and suppresses lower confidence scores in a non-linear manner for each of the second set of features for a face track with a label.

7. The method of claim 4, wherein the label with a highest combined fusion score is used as the label for the unlabeled face track.

8. The method of claim 1, further comprising:

extracting the second set of features from the set of labeled face tracks; and

generating face models using the second set of features, the face models used to correlate the second set of features from the set of labeled face tracks to the first set of features for the unlabeled face track.

9. The method of claim 1, wherein correlating comprises:

determining the feature weights for the second set of features based on a similarity of the second set of features to the first set of features.

10. The method of claim 9, wherein correlating comprises:

determining the second set of features that are nearest neighbors to the first set of features.

11. The method of claim 10, wherein each label for the set of labeled face tracks includes a respective second set of features that are nearest neighbors to the first set of features.

12. The method of claim 1, wherein the weighting function comprises:

w_j(c)=v_j1c²+v_j2c, where c² is a non-linear part that magnifies confidence scores, c is a linear part for confidence scores, and v_j are fixed parameters.

13. A non-transitory computer-readable storage medium containing instructions, that when executed, control a computer system to be configured for:

determining a plurality of face tracks for detected faces in a video;

receiving a set of labels for a portion of the plurality of face tracks from a set of users to form a set of labeled face tracks, wherein a label in the set of labels identifies an identity for a face track in the set of labeled face tracks;

extracting a first set of features for an unlabeled face track in the plurality of face tracks, wherein an identity for the unlabeled face track is not known;

correlating the extracted first set of features for the unlabeled face track to a second set of features from the set of labeled face tracks;

generating feature weights for the first set of features based on confidence scores for the second set of features using a weighting function that magnifies feature weights for the second set of features with higher confidence scores and suppresses feature weights for the second set of features with lower confidence scores in a non-linear manner; and

using the generated feature weights to determine a label for the unlabeled face track by applying the generated feature weights to the second set of features.

14. The non-transitory computer-readable storage medium of claim 13, wherein generating the feature weights comprises:

selecting the second set of features for the set of labeled face tracks;

generating a feature vector for each feature in the second set of features based on the correlating; and

generating feature weights based on the feature vector for each feature in the second set of features.

15. The non-transitory computer-readable storage medium of claim 14,

wherein the feature vector comprises a non-linear portion and a linear portion.

16. The non-transitory computer-readable storage medium of claim 13,

wherein using the generated feature weights to generate the label comprises:

combining the feature weights for the first set of features to generate combined fusion scores for each face track in the portion of the plurality of face tracks; and

determining the label based on the combined fusion scores for each face track.

17. The non-transitory computer-readable storage medium of claim 16, wherein the combined fusion score magnifies higher confidence scores and suppresses lower confidence scores in a non-linear manner for each of the second set of features for a face track with a label.

18. The non-transitory computer-readable storage medium of claim 13, further configured for:

extracting the second set of features from the set of labeled face tracks; and

generating face models using the second set of features, the face models used to correlate the second set of features from the set of labeled face tracks to the first set of features for the unlabeled face track.

19. The non-transitory computer-readable storage medium of claim 13, wherein correlating comprises:

determining the feature weights for the second set of features based on a similarity of the second set of features to the first set of features.

20. An apparatus comprising:

one or more computer processors; and

a non-transitory computer-readable storage medium comprising instructions, that when executed, control the one or more computer processors to be configured for;

determining a plurality of face tracks for detected faces in a video;

receiving a set of labels for a portion of the plurality of face tracks from a set of users to form a set of labeled face tracks, wherein a label in the set of labels identifies an identity for a face track in the set of labeled face tracks;

extracting a first set of features for an unlabeled face track in the plurality of face tracks, wherein an identity for the unlabeled face track is not known;

correlating the extracted first set of features for the unlabeled face track to a second set of features from the set of labeled face tracks;

generating feature weights for the first set of features based on confidence scores for the second set of features using a weighting function that magnifies feature weights for the second set of features with higher confidence scores and suppresses feature weights for the second set of features with lower confidence scores in a non-linear manner; and

using the generated feature weights to determine a label for the unlabeled face track by applying the generated feature weights to the second set of features.