Kurly Cuts Delivery Photo Inspection Costs by 93% with YOLOv11-Based Object Detection

Kurly's Data Service Development Team used the YOLOv11 object detection model to cut manual inspection costs for delivery-completion photos by 93%. Deliveries without purple boxes or paper bags carry a high risk of freshness degradation and require quick verification, so the team automated the process so that operators are only asked to check cases where the AI model detects neither object.

Data-Centric Approach and Labeling Efficiency

To secure training data, the team used a data engine. Instead of manually labeling every photo, they adopted an approach where only the low-accuracy portions of the model's inference results (pseudo labels) were corrected by humans. This involved a two-stage process: a result verifier performed binary classification of pseudo-label accuracy, and a detailed corrector directly fixed the bounding boxes for items classified as negative. Through this process, the rate of direct bounding box correction was kept to around 7%, minimizing labeling costs.

Model Performance and Operational Results

A review-domain model was used as the initial model, then fine-tuned for the delivery-completion domain. The purple box detection model improved after one labeling improvement cycle, reaching precision of 0.9992 and recall of 0.9792. The paper bag detection model, after initial labeling of 275 images and subsequent improvement cycles, achieved precision of 0.9694 and recall of 0.9258.

On the operational side, adding a feature to detect vinyl-wrapped deliveries reduced cases requiring manual inspection by 99.8%. Additionally, the number of vinyl-wrapped delivery cases decreased by 25% after the AI model was introduced, contributing to improved service quality. The agreement rate between delivery driver app records and model results was 91.3%, and among mismatched cases, 7.3% turned out to be more accurate from the model, demonstrating the AI's detection performance.