# US 2021/0081797 A1 · Automated selection of priors for training of detection convolutional neural networks

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0081797 A1
- Document ID: US20210081797A1
- Status: Published application
- Kind: A1
- Application number: 16/772,358
- Filed: 2018-02-05
- Published: 2021-03-18
- Earliest claimed: 2018-02-05
- Assignees: Intel Corporation
- Applicants: Intel Corporation
- Inventors: Konstantin Rodyushkin; Alexander Bovyrin
- Web page: https://patentgrep.com/patent/US20210081797A1

## CPC

- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06F 18/214: Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F 18/217: {Validation; Performance evaluation; Active pattern learning techniques}
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/09: Supervised learning
- G06V 10/454: {Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]}
- G06V 10/7553: {based on shape, e.g. active shape models [ASM]}
- G06V 10/764: using classification, e.g. of video objects
- G06V 10/82: using neural networks
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

An example apparatus for selecting priors includes a training set receiver to receive a training dataset. The apparatus includes a prior generator to generate a set of redundant priors based on the training dataset. The apparatus includes an intermediate trainer to train a detection CNN using the set of redundant priors. The apparatus includes a score and location receiver to send all training samples of the training dataset to the trained detection CNN and receive responses for all of the redundant priors in the set of redundant priors. The apparatus includes a subset selector to select a subset of the set of redundant priors based on the responses.

## Claims

### Claim at position 1

1-25. (canceled)

### Claim 26

An apparatus for selecting priors, comprising:
a training set receiver to receive a training dataset;
a prior generator to generate a set of redundant priors based on the training dataset;
an intermediate trainer to train a detection convolutional neural network (CNN) using the set of redundant priors;
a score and location receiver to send all training samples of the training dataset to the trained detection CNN and receive responses for all of the redundant priors in the set of redundant priors; and
a subset selector to select a subset of priors from the set of redundant priors based on the responses.

### Claim 27 (depends on 26)

The apparatus of claim 26, comprising a final trainer to train the detection CNN based on the selected subset of priors to generate a final detection CNN.

### Claim 28 (depends on 26)

The apparatus of claim 26, wherein the subset selector is to calculate a detection quality metric based on the subset of priors by running non-maximum suppression on responses from the subset of priors to generate final detection results and calculating the detection quality metric based on the final detection results.

### Claim 29 (depends on 28)

The apparatus of claim 28, wherein the detection quality metric comprises an average precision metric, a mean average precision, a miss rate, or any combination thereof.

### Claim 30 (depends on 26)

The apparatus of claim 26, wherein the prior generator is to define a redundant number of shapes of priors based on the training dataset and assign each shape of the redundant number of shapes to one or more feature maps to generate the set of redundant priors.

### Claim 31 (depends on 26)

The apparatus of claim 26, wherein the responses comprise class scores, bounding box locations, and size predictions.

### Claim 32 (depends on 26)

The apparatus of claim 26, wherein the subset selector is to iteratively prune redundant priors from the subset of priors based on a detection quality metric.

### Claim 33 (depends on 26)

The apparatus of claim 26, wherein the subset selector is to iteratively add additional priors to the subset of priors based on a detection quality metric.

### Claim 34 (depends on 26)

The apparatus of claim 26, wherein the subset selector is to iteratively prune and add priors from and to the subset of priors for a predefined number of times to select the subset of priors.

### Claim 35 (depends on 26)

The apparatus of claim 26, wherein the detection CNN comprises a single-shot multibox detector (SSD)-like object detector.

### Claim 36

A method for selecting priors, comprising:
receiving, via a processor, a training dataset;
generating, via the processor, a set of redundant priors based on the training dataset;
training, via the processor, a detection convolutional neural network (CNN) using the set of redundant priors;
sending, via the processor, all training samples of the training dataset to the trained detection CNN and receive generated responses for all of the redundant priors in the set of redundant priors; and
selecting, via the processor, a subset of priors from the set of redundant priors based on the generated responses.

### Claim 37 (depends on 36)

The method of claim 36, comprising training, via the processor, the detection CNN based on the selected subset of priors to generate a final detection CNN.

### Claim 38 (depends on 36)

The method of claim 36, wherein selecting the subset of priors comprises calculating a detection quality metric based on the subset of priors by running non-maximum suppression on responses from the subset of priors to generate final detection results and calculating the detection quality metric based on the final detection results.

### Claim 39 (depends on 38)

The method of claim 38, wherein the detection quality metric comprises an average precision metric a mean average precision, a miss rate, or any combination thereof.

### Claim 40 (depends on 36)

The method of claim 36, wherein generating the set of redundant priors based on the training dataset comprises defining a redundant number of shapes of priors based on the training dataset and assigning each shape of the redundant number of shapes to one or more feature maps to generate the set of redundant priors.

### Claim 41 (depends on 36)

The method of claim 36, wherein the generated responses comprise a class score, bounding box location, and size prediction for each of the redundant priors.

### Claim 42 (depends on 36)

The method of claim 36, wherein selecting the subset of priors comprises iteratively pruning redundant priors from the subset of priors based on a detection quality metric.

### Claim 43 (depends on 36)

The method of claim 36, wherein selecting the subset of priors comprises iteratively adding additional priors to the subset of priors based on a detection quality metric.

### Claim 44 (depends on 36)

The method of claim 36, wherein selecting the subset of priors comprises iteratively pruning and adding priors from and to the subset of priors for a predefined number of times.

### Claim 45 (depends on 36)

The method of claim 36, wherein selecting the subset of priors comprises iteratively pruning or adding priors to the subset of priors based on a detection quality metric to floating point operation ratio.

### Claim 46

At least one computer readable medium for selecting priors having instructions stored therein that, in response to being executed on a computing device, cause the computing device to:
receive a training dataset;
generate a set of redundant priors based on the training dataset;
train a detection convolutional neural network (CNN) using the set of redundant priors;
send all training samples of the training dataset to the trained detection CNN and receive generated class scores, bounding box locations, and size predictions for all of the redundant priors in the set of redundant priors from the trained detection CNN;
select a subset of the set of redundant priors based on the generated class scores, the bounding box locations, and the size predictions; and
train the detection CNN based on the selected subset of priors to generate a final detection CNN.

### Claim 47 (depends on 46)

The at least one computer readable medium of claim 46, comprising instructions to calculate a detection quality metric based on the subset by running non-maximum suppression on responses from the selected subset of priors to generate final detection results and calculating the detection quality metric based on the final detection results.

### Claim 48 (depends on 47)

The at least one computer readable medium of claim 47, wherein the detection quality metric comprises a mean average precision metric a mean average precision, a miss rate, or any combination thereof.

### Claim 49 (depends on 47)

The at least one computer readable medium of claim 47, comprising instructions to define a redundant number of shapes of priors based on the training dataset.

### Claim 50 (depends on 47)

The at least one computer readable medium of claim 47, comprising instructions to assign each shape of the redundant number of shapes to one or more feature maps to generate the set of redundant priors.
