The object-recognition table has to identify the objects placed on it, quickly, with someone else's hands in the frame and a venue's lighting overhead. Most of the work in getting there is repetition. Place an object, see what the model says, change something, place it again. Doing that on the finished table is slow, because every test means clearing the surface, restarting the app and waiting. So we built a separate rig. A plywood frame holds a camera at the same height as the real table, with a lamp mount above and a plain surface below. It is not presentable and was never meant to be seen. It exists so the gap between a change and an answer stays short.

What the rig ruled out

The first approach was to train on renders. We already had clean geometry for the objects, so generating labelled images looked like the cheap route. It was not. The model was confident on renders and hesitant on the real object, which had worn edges, fingerprints and reflections the renders never showed. The second thing we dropped was a fixed camera exposure. It held up in the studio and failed under warm spotlights, with people leaning over the table. The third was trusting a saved test set. Accuracy on the set stayed high while the table still stalled on objects placed at an angle or half off the edge. The set contained none of those frames, because we had captured it by putting things down neatly.

What the rig does now is mostly bookkeeping. Every session records frames with the prediction and the confidence attached, so a miss can be replayed instead of described from memory. A capture mode walks through positions, rotations and lighting states in one pass, which is how the awkward frames reach training. A viewer lists the misses grouped by object, so a pattern is visible rather than anecdotal. None of it is clever. Swapping a model or moving a threshold is now a short job. The table gets tested against the situations that break it. The rig stays in the studio and comes out whenever a new object joins the set.