Granted Patent B2

US 12,367,656 · App. 17/940,884 · Granted Jul 22, 2025

Inventors: Matthew A. Shreve (Campbell, CA); Robert R. Price (Palo Alto, CA); Jeyasri Subramanian (Sunnyvale, CA); Sumeet Menon (Baltimore, MD)

Assignee: Xerox Corporation

G06V10/764G06V10/26G06V10/7753G06V10/776G06V10/82G06V20/41

Abstract

Claims (102)

1. A computer-implemented method, comprising:

receiving an input video and a first annotated image from the input video, wherein the first annotated image identifies an object of interest in the input video;

initiating a tracker based on the first annotated image and the input video beginning from a start of the input video;

generating, by the tracker based on the first annotated image and the input video, information including: a sliding window for false positives; a first set of unlabeled images from the input video; and at least two images in which the object of interest is labeled with its corresponding state;

classifying, by a semi-supervised classifier based on the information, the first set of unlabeled images from the input video;

responsive to determining that a first unlabeled image is classified as a false positive, reinitiating the tracker based on a second annotated image and the input video beginning from a frame with the second annotated image, wherein the frame with the second annotated image occurs in the input video prior to a frame with the first unlabeled image classified as a false positive; and

generating an output video comprising the input video displayed with tracking on the object of interest, wherein the object of interest in each image from the input video is annotated and labeled with its corresponding state.

2. The method of claim 1 , wherein the first and second annotated images comprise at least one of:

a bounding box or area defined by a user; and

a bounding box or area automatically generated by a computer.

3. The method of claim 1 , wherein the at least two images in which the object of interest is labeled with its corresponding state are labeled based on at least one of:

a state defined by a user;

a state determined by the user upon a query or other prompt from a computer; and

a state determined or detected by the computer.

4. The method of claim 1 ,

wherein the object of interest in the input video comprises a region of interest in the input video.

5. The method of claim 1 ,

wherein the sliding window for false positives comprises background images corresponding to a background state, and

wherein the semi-supervised classifier classifies the first set of unlabeled images into at least one of: the background state; and a corresponding labeled state for the at least two images.

6. The method of claim 1 , wherein the semi-supervised classifier comprises at least one of:

a FixMatch algorithm;

a SimCLR framework; and

a ReMixMatch algorithm.

7. The method of claim 1 , wherein classifying the first set of unlabeled images from the input video comprises:

applying an overall loss function which is based on a supervised loss function and an unsupervised loss function,

wherein the supervised loss function is applied to the at least two images with corresponding labeled states, and

wherein the unsupervised loss function is applied to the first set of unlabeled images.

8. The method of claim 7 , wherein for a respective unlabeled image of the first set of unlabeled images from the input video, the method further comprises:

generating a weakly augmented image of the respective unlabeled image;

generating a strongly augmented image of the respective unlabeled image;

generating, by a convolutional neural network (CNN), a first prediction of a state for the weakly augmented image;

responsive to determining that the first prediction is greater than a predetermined threshold, generating, based on the first prediction, a pseudo-label for the respective unlabeled image;

generating, by the CNN, a second prediction of a state for the strongly augmented image; and

applying the unsupervised loss function to the respective unlabeled image, wherein the unsupervised loss function minimizes a loss between the first prediction and the second prediction.

9. The method of claim 1 , further comprising:

determining a third annotated image from the input video, wherein the third annotated image identifies the object of interest at an end of the input video;

initiating the tracker based on the third annotated image and the input video in reverse beginning from the end of the input video; and

generating, by the tracker based on the third annotated image and the input video in reverse, a second set of unlabeled images from the input video,

wherein the first set of unlabeled images includes the second set of unlabeled images.

10. A computer system, the system comprising:

a processor; and

a storage device storing instructions that when executed by the processor cause the processor to perform a method, the method comprising:

receiving an input video and a first annotated image from the input video, wherein the first annotated image identifies an object of interest in the input video;

initiating a tracker based on the first annotated image and the input video beginning from a start of the input video;

generating, by the tracker based on the first annotated image and the input video, information including: a sliding window for false positives; a first set of unlabeled images from the input video; and at least two images in which the object of interest is labeled with its corresponding state;

classifying, by a semi-supervised classifier based on the information, the first set of unlabeled images from the input video;

responsive to determining that a first unlabeled image is classified as a false positive, reinitiating the tracker based on a second annotated image and the input video beginning from a frame with the second annotated image, wherein the frame with the second annotated image occurs in the input video prior to a frame with the first unlabeled image classified as a false positive; and

generating an output video comprising the input video displayed with tracking on the object of interest, wherein the object of interest in each image from the input video is annotated and labeled with its corresponding state.

11. The computer system of claim 10 ,

wherein the first and second annotated images comprise at least one of:

a bounding box or area defined by a user; and

a bounding box or area automatically generated by a computer, and

wherein the object of interest in the input video comprises a region of interest in the input video.

12. The computer system of claim 10 , wherein the at least two images in which the object of interest is labeled with its corresponding state are labeled based on at least one of:

a state defined by a user;

a state determined by the user upon a query or other prompt from a computer; and

a state determined or detected by the computer.

13. The computer system of claim 10 ,

wherein the sliding window for false positives comprises background images corresponding to a background state, and

wherein the semi-supervised classifier classifies the first set of unlabeled images into at least one of: the background state; and a corresponding labeled state for the at least two images.

14. The computer system of claim 10 , wherein the semi-supervised classifier comprises at least one of:

a FixMatch algorithm;

a SimCLR framework; and

a ReMixMatch algorithm.

15. The computer system of claim 10 , wherein classifying the first set of unlabeled images from the input video comprises:

applying an overall loss function which is based on a supervised loss function and an unsupervised loss function,

wherein the supervised loss function is applied to the at least two images with corresponding labeled states, and

wherein the unsupervised loss function is applied to the first set of unlabeled images.

16. The computer system of claim 15 , wherein for a respective unlabeled image of the first set of unlabeled images from the input video, the method further comprises:

generating a weakly augmented image of the respective unlabeled image;

generating a strongly augmented image of the respective unlabeled image;

generating, by a convolutional neural network (CNN), a first prediction of a state for the weakly augmented image;

responsive to determining that the first prediction is greater than a predetermined threshold, generating, based on the first prediction, a pseudo-label for the respective unlabeled image;

generating, by the CNN, a second prediction of a state for the strongly augmented image; and

applying the unsupervised loss function to the respective unlabeled image, wherein the unsupervised loss function minimizes a loss between the first prediction and the second prediction.

17. The computer system of claim 10 , wherein the method further comprises:

determining a third annotated image from the input video, wherein the third annotated image identifies the object of interest at an end of the input video;

initiating the tracker based on the third annotated image and the input video in reverse beginning from the end of the input video; and

generating, by the tracker based on the third annotated image and the input video in reverse, a second set of unlabeled images from the input video,

wherein the first set of unlabeled images includes the second set of unlabeled images.

18. A non-transitory computer readable storage medium storing instructions which, when executed by a computer cause the computer to perform a method, the method comprising:

receiving an input video;

determining a first annotated image from the input video, wherein the first annotated image identifies an object of interest in the input video;

initiating a tracker based on the first annotated image and the input video beginning from a start of the input video;

generating, based on the first annotated image and the input video, information including: a sliding window for false positives; a first set of unlabeled images from the input video; and at least two images in which the object of interest is labeled with its corresponding state;

classifying, by a semi-supervised classifier based on the information, the first set of unlabeled images from the input video;

determining whether a first unlabeled image is classified as a false positive,

responsive to determining that a first unlabeled image is classified as a false positive, reinitiating the tracker based on a second annotated image and the input video beginning from a frame with the second annotated image, wherein the frame with the second annotated image occurs in the input video prior to a frame with the first unlabeled image classified as a false positive; and

generating an output video comprising the input video displayed with tracking on the object of interest, wherein the object of interest in each image from the input video is annotated and labeled with its corresponding state.

19. The non-transitory computer readable storage medium of claim 18 ,

wherein the first and second annotated images comprise at least one of:

a bounding box or area defined by a user; and

a bounding box or area automatically generated by a computer; and

wherein the at least two images with the corresponding labeled states are labeled based on at least one of:

a state defined by a user;

a state determined by the user upon a query or other prompt from a computer; and

a state determined or detected by the computer.

20. The non-transitory computer readable storage medium of claim 18 , wherein the method further comprises:

determining a third annotated image from the input video, wherein the third annotated image identifies the object of interest at an end of the input video;

initiating the tracker based on the third annotated image and the input video in reverse beginning from the end of the input video; and

generating, based on the third annotated image and the input video in reverse, a second set of unlabeled images from the input video,

wherein the first set of unlabeled images includes the second set of unlabeled images.

Assignments (8)

Continuity (1)

References Cited (44)