What is a backbone?
A backbone is the part of a model that does the heavy lifting of turning raw input into useful features. In visual AI it takes an image and produces feature maps that capture edges, textures, parts, and objects at multiple scales. Task-specific heads then sit on top of the backbone to produce the actual output, whether that is class labels, bounding boxes, or segmentation masks.
Because the backbone is generic, the same one can serve many tasks, which is why pretrained backbones are reused so heavily.
Key takeaways
- The backbone extracts general features that task heads build on.
- The same backbone can support classification, detection, and segmentation.
- Pretrained backbones are a common starting point via transfer learning.
How it works
An image passes through the backbone, commonly a convolutional network or a Vision Transformer, which produces feature representations at one or more resolutions. Those features feed into heads designed for the target task. Swapping the head while keeping the backbone lets a team retarget a proven feature extractor to a new problem quickly.
Why it matters
The backbone is where most of a model's transferable knowledge lives, so choosing and reusing a strong pretrained backbone is one of the fastest ways to reach good performance on a new task. This modularity, a shared backbone with interchangeable heads, is a foundational pattern across visual AI systems.
Frequently asked questions
Is a backbone the whole model?
No. It is the feature-extracting core. Task-specific heads sit on top of it to produce the final predictions.
Can the same backbone be used for different tasks?
Yes. A single backbone can feed classification, detection, or segmentation heads, which is why backbones are reused so widely.
Related terms