accuracy (backbone selection)

How well a backbone's features perform on your actual target task, not just on whatever benchmark a paper originally reported.

Accuracy, as a backbone-selection axis, measures how well a backbone's features -- plus whatever task-specific head sits on top -- perform on the actual target task, not just on whatever benchmark a paper originally reported. A backbone with a higher published benchmark accuracy is not guaranteed to be more accurate on a specific target task, especially one visually unlike the benchmark's own photo collection.

Examples

A backbone with the highest published ImageNet accuracy is not automatically the most accurate choice for classifying microscope slides, a visually different task from ImageNet's own photos.