supportgate
Guide

Product image datasets for e-commerce: quality checks before training

Published 12 October 2026

On this page

Where product images go wrong

Catalogue photos, supplier images and customer return photos come from many cameras and processes. Copies, blurry shots and mislabeled products are common, and they quietly reduce the quality of search and detection models.

What to check

  • Duplicates: the same product photo from several suppliers or in several sizes.
  • Blur and exposure: especially in customer and warehouse photos.
  • Broken files: failed uploads and unsupported formats.
  • Text and watermarks: supplier logos and price labels the model may learn by mistake.
  • People: faces and hands in returns photos are personal data; count them and decide whether you need them.

Returns photos

For damage detection, the photo must show the damage clearly. Flag photos where blur, glare or framing make the product condition impossible to judge, and send them for review instead of training on them.

Keep the decision with your team

Flag and sort rather than delete, so merchandisers and data teams can decide what counts as a usable image for each model.