RECOGNITION AND RECONSTRUCTION OF OBJECTS WITH PARTIAL APPEARANCE

• US 20190012802A1
• Filed: 07/06/2017
• Published: 01/10/2019
• Est. Priority Date: 07/06/2017
• Status: Active Grant

• Current Assignee: Futurewei Technologies Incorporated (Huawei Investment & Holding Co., Ltd.)
• Original Assignee: Futurewei Technologies Incorporated (Huawei Investment & Holding Co., Ltd.)
• Inventors: Liu, Lifeng, Yin, Xiaotian, Zhu, Yingxuan, Zhang, Jun, Li, Jian
• Granted Patent: US 10,460,470 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/24143 Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06T 2207/20084 Artificial neural networks [ANN]
• CPC Class Codes: G06T 7/73 using feature-based methods
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 20/10 Terrestrial scenes scenes under surveillance with static cameras G06V20/52; scenes perceived from the exterior of a vehicle G06V20/56; scenes perceived from the interior of a vehicle G06V20/59
• CPC Class Codes: G06V 20/70 Labelling scene content, e.g. deriving syntactic or semantic representations
• CPC Class Codes: G06V 30/19147 Obtaining sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 30/19173 Classification techniques
• CPC Class Codes: G06V 30/2504 Coarse or fine approaches, e.g. resolution of ambiguities or multiscale approaches

Abstract

Various embodiments include systems and methods structured to provide recognition of an object in an image using a learning module trained using decomposition of the object into components in a number of training images. The training can be based on an overall objectness score of the object, an objectness score of each component of the object, a pose of the object, and a pose of each component of the object for each training image input. Additional systems and methods can be implemented in a variety of applications.

Claims

1. A system operable to recognize and reconstruct objects in images, the system comprising:

a non-transitory memory storage comprising instructions; and

one or more processors in communication with the non-transitory memory storage, wherein the one or more processors execute the instructions to perform operations comprising;

acquiring a plurality of images, each image including an object within the image, the object having a different appearance in a number of images of the plurality of images;

decomposing, for each image, the object into components and generating label data for each image;

inputting the images and associated label data into a learning module to train the learning module to recognize the components of the object, the training being based on an overall objectness score of the object, an objectness score of each component of the object, a pose of the object, and a pose of each component of the object for each image input;

inputting an additional image into the trained learning module;

detecting that the additional image has one or more of the components of the object within the additional image in response to input into the trained learning module; and

identifying the object and estimating its pose information in the additional image and/or constructing a complete view of the object in the additional image or a complete sketch of the object in response to detection of the one or more of the components.

2. The system of claim 1, wherein decomposing the object in the images into components includes using a user interface to provide user input to decompose the object into the components.

3. The system of claim 2, wherein providing user input to decompose the object into components includes using one or more actions from a set of actions, the set including using heuristics to decompose the object, using a guess of a best fit from generated or matched existing decomposed candidates, and manually decomposing a number of typical examples.

4. The system of claim 1, wherein the learning module is a deep neural network.

5. The system of claim 4, wherein training the deep neural network includes

inputting the images and associated label data to a first convolution layer of a plurality of convolution layers i, i=1, . . . n−

1, n arranged in series such that input to each convolution layer includes an output of a previous convolution layer in the series beginning with convolution layer 2, convolution layer 1 being the first convolution layer;

generating, from the output of convolution layer n−

1, region proposals for each component;

generating, from the output of convolution layer n, region proposals for the object in whole; and

performing region of interest pooling using the region proposals for each component, output of convolution layers n−

1 and i, and the region proposals for the object in whole.

6. The system of claim 5, wherein the region of interest pooling includes a consistency checking among the region proposals for the object in whole and the region proposals for the components to discard inconsistent region proposals.

7. The system of claim 1, wherein training the learning module includes minimizing a cost function generated as a sum of discrepancies between the label data and computed data for a set of items, the set of items including the overall objectness score, the objectness scores of the components, a pose of the object, and the poses of the components.

8. The system of claim 1, wherein the plurality of images includes one or more of a full appearance of the object, a partial appearance of the object, an orientation variation of the object, or an appearance variation of the object.

9. The system of claim 1, wherein acquiring the plurality of images includes creating training images of the object with partial appearance of the object in the training images and labeling the training images with pose information from creating the training images or labeling the training images with pose information based on deformation descriptors from conducting an object template matching.

10. The system of claim 1, wherein constructing the complete view includes applying pose information from detecting that the additional image has one or more of the components of the object within the additional image and the additional image as input to a generative adversarial network to reconstruct a full appearance of the object.

11. A computer-implemented method comprising:

acquiring a plurality of images, each image including an object within the image, the object having a different appearance in a number of images of the plurality of images;

decomposing, for each image, the object into components and generating label data for each image;

inputting the images and associated label data into a learning module to train the learning module to recognize the components of the object, the training being based on an overall objectness score of the object, an objectness score of each component of the object, a pose of the object, and a pose of each component of the object for each image input;

inputting an additional image into the trained learning module;

detecting that the additional image has one or more of the components of the object within the additional image in response to input into the trained learning module; and

identifying the object and estimating its pose information in the additional image and/or constructing a complete view of the object in the additional image or a complete sketch of the object in response to detection of the one or more of the components.

12. The computer-implemented method of claim 11, wherein decomposing the object in the images into components includes using a user interface to provide user input to decompose the object into the components.

13. The computer-implemented method of claim 12, wherein providing user input to decompose the object into components includes using one or more actions from a set of actions, the set including using heuristics to decompose the object, using a guess of a best fit from generated or matched existing decomposed candidates, and manually decomposing a number of typical examples.

14. The computer-implemented method of claim 11, wherein the learning module is a deep neural network.

15. The computer-implemented method of claim 14, wherein training the deep neural network includes

inputting the images and associated label data to a first convolution layer of a plurality of convolution layers i, i=1, . . . n−

1, n arranged in series such that input to each convolution layer includes an output of a previous convolution layer in the series beginning with convolution layer 2, convolution layer 1 being the first convolution layer;

generating, from the output of convolution layer n−

1, region proposals for each component;

generating, from the output of convolution layer n, region proposals for the object in whole; and

performing region of interest pooling using the region proposals for each component, output of convolution layers n−

1 and n, and the region proposals for the object in whole.

16. The computer-implemented method of claim 11, wherein training the learning module includes minimizing a cost function generated as a sum of discrepancies between the label data and computed data for a set of items, the set of items including the overall objectness score, the objectness scores of the components, a pose of the object, and the poses of the components.

17. The computer-implemented method of claim 11, wherein acquiring the plurality of images includes acquiring one or more of a full appearance of the object, a partial appearance of the object, an orientation variation of the object, or an appearance variation of the object.

18. The computer-implemented method of claim 11, wherein acquiring the plurality of images includes creating training images of the object with partial appearance of the object in the training images and labeling the training images with pose information from creating the training images or labeling the training images with pose information based on deformation descriptors from conducting an object template matching.

19. The computer-implemented method of claim 11, wherein constructing the complete view includes making a correspondence between an object template and a detected object having a partial appearance.

20. The computer-implemented method of claim 11, wherein constructing the complete view includes applying pose information from detecting that the additional image has one or more of the components of the object within the additional image and the additional image as input to a generative adversarial network to construct a full appearance of the object.