Why generic vision breaks
A model that nails the demo can miss on your site.
Computer vision research has a name for it: domain shift. A model trained in one environment loses accuracy the moment the environment changes. Take a detector tuned on bright, clean images and drop it into a mine at depth, sodium-vapor light, heavy dust, reflective rock, narrow tunnels, camera distortion, and detection degrades fast. The same gap shows up in luxury, where a generic image model trained on the whole internet cannot hold a brand's exact stone count, proportions, or hardware placement. The fix is adaptation: fine-tune the model on your environment, your cameras, and your rules, and it starts to earn its place. Generic vision tools are built for clean, predictable environments, and mines, plants, and high-mix production sites are none of those.