The Model Is the Easy Part: What a Real-Time Computer Vision Product Actually Takes

Short summary

An experienced practitioner argues that real-time computer vision is a pipeline problem, not a model problem. The model is one node in a sense-decide-act loop where latency, edge-vs-cloud tradeoffs, and adversarial real-world conditions determine success. Using a real product (Raqts, a responsive racquet-sport wall) as a case study, the author explains why demo accuracy and production accuracy are different numbers.

• •Computer vision products are real-time loops, not just models — the pipeline is the hard part
• •Production accuracy differs from demo accuracy due to adversarial real-world conditions
• •End-to-end latency budget (camera to inference to action) is a hard constraint that shapes all other decisions

Generated with AI, which can make mistakes.