The core trade-off: accuracy against frames per second. Each point is a model, coloured by family and sized by input resolution. Turn on the Pareto frontier to see the models that are not beaten on both axes at once. Accuracy is only comparable within one dataset at one input size, so the view is always scoped to both.