Kolon Kurly Automates Review Image Classification and Builds Training Data with AWS SageMaker Ground Truth

Market Kurly operates a machine learning model to automate the classification of customer review images, and within one year of adoption, reduced the proportion of screenshot reviews from about 10% to the 1% range. However, as the model ran without additional training, demands to improve misclassifications emerged, prompting a review of the training data preparation process to address this.

Difficulties in Training Data Preparation and Initial Methods

Review images are stored in S3 by date, with 6,000 to 8,000 uploaded daily. Initially, hashlib was used to remove duplicate images, and PIL was used to filter out corrupted images. For classification, drag-and-drop using Finder's preview feature (about 1 hour per 10,000 images) and a command-line classifier (about 1 hour 10 minutes per 10,000 images) were used, but as the data volume grew to the scale of 100,000 images, this became inefficient due to issues such as Finder freezing.

Adoption of AWS SageMaker Ground Truth and Cost Analysis

As an alternative, AWS SageMaker Ground Truth was reviewed, and since the validity of automatic classification algorithms was judged to be low, a manual classification approach using designated personnel was chosen. Ground Truth supports manifest file creation and worker screen configuration, and through its private tenant option, internal personnel or external vendors can be utilized.

However, limitations emerged in terms of cost. When outsourcing classification through AWS, about $1,200 (about 1.5 million won) was expected for 100,000 images, whereas processing 100,000 images with domestic personnel (assuming an hourly wage of 10,000 won) would require about 6 million won, making Ground Truth appear cost-effective. However, Ground Truth's private labeling also incurred a fee of about $400 (about 500,000 won) per 100,000 items, resulting in a still-significant cost burden for large-scale data processing. In the end, despite the convenience of Ground Truth, the cost issue means alternatives need to be explored.