Building Real-Time Computer Vision Products Is a Pipeline Problem, Not a Model Problem

Developers often assume computer vision success hinges on training an accurate model, but experienced engineers argue the real challenge lies in building a reliable end-to-end pipeline. A vision system must complete a continuous loop — sensing, deciding, and acting — fast enough that the response feels connected to the triggering event. Production environments introduce unpredictable lighting, camera angles, and scenarios absent from training data, making controlled-test accuracy a poor predictor of real-world performance. Latency is treated as a hard engineering constraint, similar to frame budgets in game development, meaning a slow but accurate system fails just as badly as an inaccurate one. The author draws on firsthand experience developing Raqts, a real-time reactive racquet-sport wall, to illustrate how any delay in the vision pipeline makes the entire product feel broken rather than merely imprecise.