CODEWITHSUNDEEP
machine-learningcomputer-visionimage-segmentationsemantic-segmentation
world4jason
world4jason

Reputation: 11

Should I subtract image mean when doing a task of semantic segmentation? why or why not?

Should I subtract image mean or divide 255 for a image on a semantic segmentation task? why or why not?

Upvotes: 1

Views: 99

Answers (1)

Shai
Shai

Reputation: 114926

You should.

The way filter weights are initialized there is a hidden assumption that the input signal is roughly zero mean and unit variance. This helps scaling the activations and the gradients.
In practice, you may be able to train a net (for segmentation or any other task for that matter) using the "un-normalized" input images, but it will probably take longer to train and be less stable w.r.t meta parameters such as learning rate and solver type.

For more details see section 2.2 in He et al. "Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification" (ICCV 2015).

Upvotes: 0

Related Questions