Granted Patent B2

US 12,561,964 · App. 18/023,506 · Granted Feb 24, 2026

Inventors: Changdong Yoo (Daejeon, KR); Thang Vu (Daejeon, KR); Xuan Trung Pham (Daejeon, KR); Hyunjun Jang (Daejeon, KR)

Assignee: KOREA ADVANCED INSTITUTE OF SCIENCE AND TECHNOLOGY

G06V10/82G06V10/776

Abstract

Claims (61)

1 . A method comprising:

receiving an input image and generating a first convolutional feature map;

converting the first convolutional feature map into a second convolutional feature map by a first convolution;

generating a first anchor for each point of the input image;

determining whether the first anchor is within a ground truth box;

generating a second anchor on the second convolutional feature map by scaling and shifting the first anchor within the ground truth box;

generating a third convolutional feature map by convolving the second convolutional feature map by a second convolution;

determining whether an overlap ratio between the ground truth box and the second single anchor is greater than or equal to a reference value;

generating a third anchor by scaling and shifting the second anchor having an overlap ratio greater than or equal to the reference value;

assigning an objectness score to the third anchor;

proposing the third anchor having an objectness score equal to or greater than a reference value as a proposal on the third convolutional feature map;

generating a fourth convolutional feature map by convolving the third convolutional feature map by a third convolution;

determining whether an overlapping ratio between the ground truth box and the third single anchor is greater than or equal to a reference value;

generating a fourth anchor by scaling and shifting the third anchor having the overlapping ratio greater than or equal to the reference value;

assigning an objectness score to the fourth anchor; and

proposing the fourth anchor having the objectness score equal to or greater than a reference value as a proposal on the fourth convolutional feature map.

2 . The method of claim 1 , further comprising identifying a candidate object at the point based on the objectness score.

3 . The method of claim 2 , further comprising:

determining a category of the candidate object; and

assigning a confidence score to the category of the candidate object.

4 . The method of claim 1 , wherein the first convolution is a dilated convolution.

5 . The method of claim 1 , wherein the second convolution is an adaptive convolution.

6 . A system comprising:

a processor; and

a computer-readable medium including instructions for executing an object detection and classification network by the processor,

wherein the object detection and classification network includes an initial processing module configured to input an image and generate a convolutional feature map, and an object proposal module configured to generate a proposal corresponding to a candidate object in the image, and

wherein the object proposal module performs:

converting a first convolutional feature map into a second convolutional feature map by a first convolution;

generating a first anchor for each point of the input image;

determining whether the first anchor is within a ground truth box;

generating a second anchor on the second convolutional feature map by scaling and shifting the first anchor within the ground truth box;

generating a third convolutional feature map by convolving the second convolutional feature map by a second convolution;

determining whether an overlap ratio between the ground truth box and the second single anchor is greater than or equal to a reference value;

generating a third anchor by scaling and shifting the second anchor having an overlap ratio greater than or equal to the reference value;

assigning an objectness score to the third anchor;

proposing the third anchor having an objectness score equal to or greater than a reference value as a proposal on the third convolutional feature map;

generating a fourth convolutional feature map by convolving the third convolutional feature map by a third convolution;

determining whether an overlapping ratio between the ground truth box and the third single anchor is greater than or equal to a reference value;

generating a fourth anchor by scaling and shifting the third anchor having the overlapping ratio greater than or equal to the reference value;

assigning an objectness score to the fourth anchor; and

proposing the fourth anchor having the objectness score equal to or greater than a reference value as a proposal on the fourth convolutional feature map.

7 . The system of claim 6 , wherein the first convolution is a dilated convolution.

8 . The system of claim 6 , wherein the second convolution is an adaptive convolution.

9 . The system of claim 6 , wherein the object proposal module performs identifying the candidate object at the point based on the objectness score.

10 . The system of claim 6 , further comprising a proposal classifier for determining a category of the candidate object and assigning a confidence score to the category of the candidate object.

11 . A system comprising:

a processor; and

a computer-readable medium including instructions for executing an object detection and classification network by the processor,

wherein the object detection and classification network includes an initial processing module configured to input an image and generate a convolutional feature map, and an object proposal module configured to generate a proposal corresponding to a candidate object in the image, and

wherein the object proposal module performs:

converting a first convolutional feature map into a second convolutional feature map by a first convolution;

generating a first anchor for each point of the input image;

determining whether the first anchor is within a ground truth box;

generating a second anchor on the second convolutional feature map by scaling and shifting the first anchor within the ground truth box;

generating a third convolutional feature map by convolving the second convolutional feature map by a second convolution;

determining whether an overlap ratio between the ground truth box and the second single anchor is greater than or equal to a reference value;

generating a third anchor by scaling and shifting the second anchor having an overlap ratio greater than or equal to the reference value;

assigning an objectness score to the third anchor; and

proposing the third anchor having an objectness score equal to or greater than a reference value as a proposal on the third convolutional feature map,

further comprising a proposal classifier for determining a category of the candidate object and assigning a confidence score to the category of the candidate object,

further comprising a machine learning module for training at least one parameter of the initial processing module and the object proposal module to generate at least one proposal on a training image, and training at least one parameter of the proposal classifier module to assign a category to each of the at least one proposal on the training image.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (26)