backbone (vision model)

The shared feature-extraction network that one or more task-specific heads are attached to.

A backbone is the shared feature-extraction portion of a vision model -- a CNN or Vision Transformer that turns an input image into a feature representation -- which one or more task-specific heads (a classifier, a detector, a segmenter) are then attached to. A single pretrained backbone is often reused across many downstream tasks via transfer learning.

Examples

The same pretrained ResNet backbone feeds a classification head in one project and an object-detection head in another, with no retraining of the backbone itself needed for either.