Reducing Manual Review Costs by Detecting Objects in Delivery Completion Photos

Project publication date, project published by Kurly ML engineer Yoon Junho on 2025-10-27.

Project goal automate detection to reduce manual review, goal is to detect purple boxes and paper bags in delivery completion photos to reduce manual reviews.

Data engine applied data-centric labeling and retraining loop, used a data-centric workflow: pseudo-labeling unlabeled images, human verification/correction, and periodic retraining.

Labeling process used two-role human labeling process, applied two human roles for label quality: Result Validator (binary check) and Detail Corrector (bounding box edits).

Model and tooling choice, object detection model used YOLOv11 with Ultralytics tooling.