Best Real-Time Vision Models

The same architecture ships in sizes from nano to extra-large, and the right size is set by your hardware and frame rate, not by peak benchmark accuracy. A model that hits 80 FPS on a data-center GPU may run at 4 FPS on the edge box you can afford to put on site. Fix the target device and required FPS first, then choose the largest model that fits with headroom; published FPS numbers are upper bounds measured on strong GPUs.

Input resolution is the strongest lever in the family: halving it can double throughput, at the price of missing small objects. Detection-plus-tracking often beats detecting on every frame (detect every few frames, track between them). And not every pipeline needs real time at all; if the video is recorded and analyzed later, batch processing with a bigger, more accurate model is usually the better trade.

A compact model fine-tuned on your own footage routinely beats a larger generic one on both accuracy and speed for a fixed camera and class list. Before stepping up a size (and losing your latency headroom), try training the size you have on data from the actual deployment.