Skip to content
Quantoros

What does an auxiliary segmentation head contribute during training and inference?

Computer VisionLevel: IntermediateDeep supervision versus inference heads

What does an auxiliary segmentation head contribute during training and inference?

Hints
  • Separate a training loss path from the prediction path.
  • Does the final mask combine both heads in this architecture?

Answer

It applies an additional loss to intermediate features, providing deep supervision during training. In the standard setup described by the FAQ, final predictions come from the decode head and the auxiliary head is not an inference ensemble. Check loss weights and label handling for both heads; extra supervision can affect learned features even when its head is removed at inference.

Source and licence: OpenMMLab — MMSegmentation FAQ · Apache-2.0 · answer: Quantoros editorial

Related questions