MalaysiaRetail and e-commerce2021 – ongoing

Carsome

Automated vehicle-number-plate masking

Carsome reported masking car plates in 2.4 seconds per vehicle with computer vision

WorkingCompany's own claimOfficial ↗

Problem

Inspectors manually blurred number plates in vehicle photographs to protect customer privacy. The task took at least 1 minute per car.

Approach

Carsome built a Mask R-CNN model after finding that YOLOv3 produced poorly fitted masking regions. It ran scheduled image batches through Amazon SageMaker Batch Transform, coordinated with Apache Airflow. Masked images were stored for subsequent retrieval.

Results

2.4 secondsautomated masking time per car, versus at least 1 minute manually, according to CarsomeCompany's own claim[1]

99correctly blurred images in a test of 100 Malaysian car imagesCompany's own claim[1]

Takeaway

Scheduled batch processing suited a privacy task that did not require immediate results.

Sources

CARSOME: Leveraging Automatic Car Plate Masking on Amazon SageMaker to Focus on Growth
Carsome21 Sep 2021Company publication

Source summaries

[1]Carsome21 Sep 2021

Carsome engineers described their deployed image-masking workflow. They reported 2.4 seconds per car and 99 correctly blurred images in a 100-image test.

Full details are in the original sources.

Added 1 Nov 2026 · Updated 1 Nov 2026