Camera Angles That Break Warehouse Computer Vision

Published 2026-07-22 by Faraz Rahimi

The model you trained on front-facing catalog shots will fail on a top-down bin photo. Angle is a first-class failure mode.

Camera Angles That Break Warehouse Computer Vision

Tags: Computer Vision, Mobile, OpenCV, Warehouse, unlisted

Operators photograph from wherever their arm already is. That is often above the bin, from the hip, or with the distinctive label facing the rack, not the camera.

If your library is all hero shots, you have trained a museum guide, not a floor tool.

Prompt for a usable frame without sounding like a designer

I use a lightweight silhouette overlay: fill the box, keep the item inside, avoid the ceiling lights. When the item is too small in frame, I ask them to step in, not "improve composition."

Top-down photos of items in bins are a separate class. I keep a few of those per SKU because that is how people actually search when the item is lying down.

Reject the unfixable

A photo of five items in a bin with no crop is not a query. Detect multiple large objects and ask which one. Guessing is how you teach the model the wrong lesson.

Original post: https://farazrahimi.com/posts/camera-angles-that-break-warehouse-computer-vision