HOW IT WORKS
Behind the Curtain
Engineered to support large scale matching of candidate images to SKUs
Background
The problem: The Motion.com website offers millions of different industrial products to businesses across North America. However, out of 12 million total SKUs, only 1 million currently have at least one product image.
Our objective: To design an automated process for sourcing missing product images for millions of SKUs, boosting sales and elevating the customer experience.
Requirements: Robust web scraper to find and download potential matches. Machine Learning model to evaluate image quality and relevance. Intuitive UI for Motion employees to review scraped images.
Core Systems
Web Scraper
Searches for product images on OEM websites, distributor websites, and Bing. Downloads images and automatically discards low quality results.
Machine Learning Model
Trained on thousands of examples. Analyzes file names and calculates image features such as brightness, whitespace, and entropy. Assigns confidence scores to each image to help prioritize human review efforts.
Review UI
Accessible by clicking "Catalog Navigator" in the navbar. Supports sorting and filtering images by manufacturer, SKU, confidence score, and date added. Reviewers' approvals and rejections are confirmed by admins.
Feedback Loop
Every approval and rejection is recorded and used to continually retrain and refine the ML Model.