Md. Asif Uddin

Proposition 1III.6.P0166 of 76 in the corpus

The task lives in the head and the loss, not in the backbone.

One encoder serves classification, detection, segmentation, depth and retrieval. What changes is the shape of the output and the function that scores it.

One image, five tasksA single image and backbone feeding five different heads: classification, detection, segmentation, depth and retrieval. The backbone is the same in each case; what differs is the head and the shape of what comes out.one imageone backbonefive headsencoderunchangedclassificationone labeldetectionboxes + labelssegmentationa label per pixeldeptha distance per pixelretrievala vector to compareNothing about the pixels changed between these five rows. What changed is the question, and with itthe output and the loss that scores it — which is why a good backbone is worth more than a good head.Classification throws the spatial axes away. Everything below it has to keep them, or get them back.
Fig. 1 — One image and one backbone feeding five heads. What changes is the question, the output shape and the loss.

Demonstration

Five questions of one photograph:

classification   image → one label
detection        image → a set of (box, label, score)
segmentation     image → a label per pixel
depth            image → a distance per pixel
retrieval        image → a vector, to compare with others

The encoder can be identical across all five. What differs is the head — a few layers — and the loss.

The organising distinction is what happens to the spatial axes.

Classification collapses them. Global average pooling turns H × W × C into C, and after that no spatial question can be answered.

Detection keeps them coarsely. Predictions are made at feature-map positions, so output stride sets the smallest object that can be localised.

Segmentation and depth must keep them fully, which is why both use encoder–decoder shapes and why Proposition 3 is about getting the resolution back.

Retrieval collapses them like classification but scores differently: there is no fixed label set, only a metric.

This is the practical content of transfer learning. A backbone pretrained for classification is useful for detection because the features are about content, and content is what all five questions are asked about. Only the last part is task-specific, which is exactly the ordering Chapter III found in the hierarchy.

Corollary

When a vision result improves, ask which of the three moved: backbone, head or loss. Papers frequently change all three and attribute the gain to the one in the title, and the ablation that separates them is the part worth reading.

Sources

Depends on

Used by