Auto Secure LoginProduct Vision / Research
Product Vision: stop losing flat products in the object count
A fresh counting study addressed product photographs that were incorrectly reported as containing no object.
Work covered: Oct 3, 2026
When the print is mistaken for the product
Flat products create a particular counting problem. A card or bookmark can fill the frame while its printed picture attracts more attention than the physical item. Earlier Product Vision readings sometimes agreed on no object at all in those photographs. The result was a zero count even though the image was a merchant’s product photograph.
A change tested on new photographs
The October 3 study introduced a minimum count of one for that specific product-photo setting. The rule and acceptance criteria were set before a fresh sample of 51 photographs was evaluated. Exact counts improved from 76.5% to 84.3% in that sample, and 96.1% were within one of the reference count. The change did not turn a previously correct count into an incorrect one in the measured set.
Do not adopt every extra check
A separate physical-piece check was also tried. It could mistake touching objects for a single piece and did not produce a net improvement on the fresh sample, so it was not adopted. That is an important outcome: another layer of analysis is useful only if it improves the actual task, not simply because it sounds more independent.
The scope of the result
These percentages describe a bounded research sample with a session-created reference, not universal counting accuracy. The minimum-one assumption belongs to merchant product photographs, not arbitrary empty images. Within that scope, the milestone fixes a concrete failure mode and gives the description workflow a more reliable inventory to start from.
Explore the product
Based on dated development and release records from September 25–October 8, 2026. Availability and acceptance are stated for the recorded milestone.