In everyday scenes, a mirror can make one object look like several, or a glass door can leave an object visible but still physically present. Counting such images is tricky because seeing every occurrence does not tell you how many real things are there. Most visual counting tests ignore this difference and just count what is detected.
The authors built PhysCount, a with 546 real-world images and 4,595 questions organized so each scene tests both counting the visible appearances and counting the physical entities that produce them. The questions also check whether models can match appearances to the same object across reflections, glass, and other viewing conditions.
The main finding is that models are far better at counting appearances than at connecting them to physical objects. Even the strongest model answered every question correctly in only 13.36% of the groups, and performance dropped sharply when one object appeared many times or through different materials. This gap matters for any AI that must understand scenes physically, such as robots or augmented reality, because a wrong count from a mirror could cause real errors.