01 / Recognize
What is actually visible?
Identify the intended object in an authorized photo or video, including distinguishing details and the limits of the view. A correct category is not a verified product identity.
The Test Bench / Case study 02 / Product retrieval
Product Match
Bench.
From an image or sampled video to candidate products—with a clear boundary between a promising search result and a verified exact match.
PUBLIC CASE STUDY · The API-metered research application is not offered as an unrestricted public launch. No paid search request is made by this page.
Question / retrieval QA
Same product—or just similar?
A plausible image match can still be the wrong manufacturer, model, part, size, or fit. The evaluation has to keep retrieval, recognition, and human verification separate.
02 / Localize
Did we crop the right thing?
Check that the search region includes the meaningful features of the selected item while excluding distracting background objects. Recognition and localization can fail independently.
03 / Retrieve
Find candidates, not verdicts.
Search using text and visual references, treating listings and image-source leads as unverified candidate evidence rather than ground truth.
04 / Review
Exact versus substitute.
A reviewer decides whether evidence supports an exact item, a potentially usable alternative, or no verified match. Functional interchangeability may require additional specifications and safety checks.
Where the prototype succeeded—and where it failed.
Synthetic two-burner stove: Product-category recognition and retrieval produced eight Google Shopping leads and six Google Lens visual references. None was independently verified as an exact match. Useful candidates are not confirmed identities.
Synthetic red-shade lamp: The model recognized a lamp with a ribbed wooden base, but its initial crop cut off distinctive portions of the red shade. A subsequent crop correction and regression tests were reported; that does not establish general real-world crop accuracy.
Focused-object search: A user-selected image region narrows analysis when a crowded scene contains unrelated products. The target is to reduce irrelevant retrieval without requiring the user to manually reverse-engineer the system.
Recognizing the object correctly does not establish that the search crop is correct. Returning a close-looking listing does not establish that the product is the same.
MULTIMODAL SEARCHLOCALIZATION QAEXACT VS SUBSTITUTEREVIEWER VERIFICATION
Access / evidence limits
What this page does not claim.
This is a public description of experimental retrieval and evaluation work, not a working copy of the paid application.
Retrieval is not validation.
Listings can be irrelevant, outdated, or misleading; exact identity needs independent corroboration. A substitute needs a separate functional assessment, not just similar appearance.
Access is deliberately limited.
The research backend uses metered search providers. A public direct-use link is withheld until authentication, request limits, and spending controls are separately reviewed. This page does not call those services.