Method, system and non-transitory computer-readable recording medium for creating training datasets for model training

1 . A method of creating a training dataset for model training, comprising the steps of:

receiving image data including at least one fashion item;

performing location box labeling on an item location box which indicates a location of an item included in the image data by using an item location detection model, thereby calculating a location box labeling result value and a location box labeling confidence value;

maintaining the location box labeling result value or changing the location box labeling result value to a null value, based on the and a location box labeling confidence value and a predetermined criterion;

receiving a user's location box review value for the maintained or changed location box labeling result value;

determining an item location box prediction value for the image data by using the item location detection model;

determining a location noise value indicating an accuracy of the location box review value by comparing the location box review value and the item location box prediction value;

determining the location box review value as location box training data if the location noise value meets a predetermined first criterion; and

creating the training dataset by associating the location box training data with the image data.

2 . The method of claim 1 , further comprising the steps of:

performing attribute labeling for an attribute of an item located within the item location box by using an item attribute recognition model;

calculating an attribute labeling result value and an attribute labeling confidence value;

receiving a user's attribute review value for the attribute labeling result value;

determining an attribute noise value of the attribute review value by using the item attribute recognition model; and

determining the attribute review value as attribute training data if the attribute noise value meets a predetermined second criterion.

3 . The method of claim 2 , further comprising: further comprising the step of:

receiving the user's attribute review value again if the attribute noise value does not meet the predetermined second criterion.

4 . The method of claim 2 , wherein, in the step of calculating the attribute labeling result value and the attribute labeling confidence value, if the attribute labeling confidence value does not meet a predetermined fourth criterion, the attribute labeling result value is determined as a null value.

5 . The method of claim 2 , further comprising the step of:

training the item attribute recognition model with the attribute training data.

6 . The method of claim 2 , wherein, in the step of receiving the user's location box review value for the location box labeling result value, the location box review value with a low location box labeling confidence value is preferentially received, and

wherein, in the step of receiving the user's attribute review value for the attribute labeling result value, the attribute review value with a low attribute labeling confidence value is preferentially received.

7 . The method of claim 1 , wherein, in the step of performing the location box labeling, the item location box is determined such that an item included in the image data is located within the item location box by using the item location detection model, and

in the step of calculating the location box labeling result value and location box labeling confidence value, a coordinate value of the item location box on the image data is determined by using the location box labeling result value, and the location box labeling confidence value indicating an accuracy of the item location box is determined.

8 . The method of claim 1 , further comprising the step of:

receiving the user's location box review value again if the location noise value does not meet the predetermined first criterion.

9 . The method of claim 1 , wherein, in the step of maintaining or changing the location box labeling result value, if the location box labeling confidence value does not meet a predetermined third criterion, the location box labeling result value is determined as the null value.

10 . The method of claim 1 , further comprising the step of:

training the item location detection model with the location box training data.

11 . A non-transitory computer-readable recording medium stored thereon a computer program for performing a method of claim 1 .

12 . A training data creating system for a model training, comprising:

an image data receiving unit configured to receive image data including at least one fashion item;

a pseudo labeling performing unit configured to perform location box labeling for an item location box which indicates a location of an item included in the image data by using an item location detection model, calculate a location box labeling result value and a location box labeling confidence value, and maintain the location box labeling result value or change the location box labeling result value to a null value based on the location box labeling confidence value and a predetermined criterion;

a labeling review value receiving unit configured to receive a user's location box review value for the maintained or changed location box labeling result value;

a noise determination unit configured to determine an item location box prediction value for the image data by using the item location detection model, and determine a location noise value indicating an accuracy of the location box review value by comparing the location box review value and the item location box prediction value; and

a training data determination unit configured to determine the location box review value as location box training data if the location noise value meets a predetermined first criterion, and create a training dataset by associating the location box training data with the image data.

13 . The training data creating system of claim 12 , wherein the pseudo labeling performing unit is configured to further perform attribute labeling for an attribute of an item located in the item location box by using an item attribute recognition model, and calculate an attribute labeling result value and an attribute labeling confidence value;

wherein the labeling review value receiving unit is configured to receive a user's attribute review value for the attribute labeling result value;

wherein the noise determination unit is configured to determine an attribute noise value of the attribute review value by using the item attribute recognition model; and

wherein the training data determination unit is configured to determine the attribute review value as attribute training data if the attribute noise value meets a predetermined second criterion.