search

ducha-aiki / affnet

Code and weights for local feature affine shape estimation paper "Repeatability Is Not Enough: Learning Discriminative Affine Regions via Discriminability"
MIT License
266 stars 47 forks source link

OriNet #12

Closed appleleaves closed 6 years ago

appleleaves commented 6 years ago

Why did not you talk about OriNet in the paper. How did you use it in the Table 1 and 2?

appleleaves commented 6 years ago

About the Table 2, I would like to ask a few questions.

  1. How do you train separately in the setting (1) ?
  2. In the setting (2), why don't you compare the biases init 0?
ducha-aiki commented 6 years ago

OriNet actually is not interesting to me by itself, because it has results similar to Yi et.al, 2015. You can seen from bottom part of Table 1 that all the variants of OriNet perform essentially the same.

And it is used for Table 2, "separately" column. Separately means that I have trained OriNet with https://github.com/ducha-aiki/affnet/blob/master/train_OriNet_test_on_graffity.py and AffNet with https://github.com/ducha-aiki/affnet/blob/master/train_AffNet_test_on_graffity.py

And combined them only in full pipeline (detection - affine - ori - desc) in test time.

In the setting (2), why don't you compare the biases init 0?

Because it failed in (1), so I thought that it is not worth it. But probably you are right and I should have try to train it anyway

appleleaves commented 6 years ago

But the parameter in setting (1,2) include ori information. How can you train "A" and keep the ori unchanged?

ducha-aiki commented 6 years ago

By applying https://github.com/ducha-aiki/affnet/blob/master/LAF.py#L279 transformation. It cancels the rotation