Skip to main content
All articles
Discovery and availability

Inside Youzu Lens: why visual search starts with the crop

How Youzu tested Florence and YOLO for object detection, why segmentation matters, and how to measure detection separately from a shopper’s total wait.

Published by Youzu · Editorial policy

Published
Updated
Reading time
3 min read

What changed: Revised article and supporting media

In this article

A shopper taps a chair in a room photo. Before Youzu Lens can search the catalog, it needs a useful picture of that chair. The engineering change described here was to object detection: in the team’s tests, a newer YOLO model was faster at this step than the tuned Florence setup. That result says nothing by itself about how long a shopper waits for results.

A patio with a table and chairs alongside a diagram of objects linked to products.
One scene contains several possible searches. The object selection matters before a result can be evaluated.

Why the team tried Florence, then returned to YOLO

An earlier YOLO setup struggled with furniture and home decor. Florence improved detection quality on those images, but the detector took too long for an interactive search. The team tried compilation, GPU changes and kernel tuning before testing a newer YOLO model. That model performed the detection step faster even before the team exported it to TensorRT and recompiled it.

The useful comparison is between those particular configurations on those images. It is not a general ranking of Florence and YOLO. A fast detector that picks the wrong object has made the wrong search faster.

Why not just draw a rectangle around the chair?

A bounding box around a chair can also contain the table behind its legs and part of a patterned rug. Segmentation identifies the pixels belonging to the selected object. The resulting image is a better input for an embedding: the numerical representation the system uses to look for visually similar catalog items. If a very tight crop cuts off the arms or the base, though, it can discard exactly the detail that sets the chair apart.

The original August 2025 engineering account also described the catalog side. In that design, product metadata arrives through SQS; a separate queue handles images. The system stores image embeddings in Qdrant, compares them using cosine similarity, and caches commonly requested groups of results. Replacing the detector doesn’t remove the time spent on these other steps.

  1. Pick a room photo with two plausible targets. Ask testers to select the same item and save the actual crop. Can a person recognize the intended product from that crop alone?
  2. Run the same catalog search with the full scene and the selected object. Inspect the first results for category, shape and identity. Record a visually similar item as similar, even if it makes a good substitute.
  3. Time detection, retrieval and the full path to visible results separately. Repeat with a cold cache and a repeated query. Include slow and failed requests rather than averaging only the successful clicks.

The Lens film shows object selection and product results; it is not a timing benchmark. For a buying decision, use the visual search evaluation guide and repeat the test with your own assortment.

Select an object and search the catalog · Watch on its own page · Open MP4
Read the captions

Youzu Lens Search your catalog with a picture. Pick an item in a photo. See matching products from your catalog. It works from your own product images. See it with your catalog. Book a demo

Next step

Automate Commerce. Make shopping personal.

Bring your catalog, listing flow or shopping idea. Let's find the right place to start.

Book a demo

Tell us about your catalog and we will get back to you.