Mask in losses.py

[Bing]

Hi Zixin,

I noticed that you apply a mask (line.134 of losses.py) with a threshold of 0.008. May I know the rules to select this value?

Best,
Bing

[Jiahui Zhang]

Hi Bing,

This threshold is used to filter too hard negative samples, or to mitigate the noise (incorrect gt correspondences) in the training data, which might harm the performance. This is also used in hardnet.

[Bing]

Hi Jiahui,

Many thanks for your kind reply. To safely remove such samples, "dist_mat_without_min_on_diag" (line 136) should plus "mask*10" as HardNet does instead of "mask". Have I got that right?

Best,
Bing

[Jiahui Zhang]

For normalized features, the maximum of dist mat is 1.0, so it is ok to plus "mask". "+mask*10" is indeed a more safe way.

[Bing]

Hi Jiahui,

If ‖𝑥‖=‖𝑦‖=1, the triangle inequality gives ‖𝑥−𝑦‖≤‖𝑥‖+‖𝑦‖=2. Thus, the values in dist_mat should range from 0.0 to 2.0. It might not be safe to just plus "mask".

Best,
Bing

[Jiahui Zhang]

Sorry, you are right. But I think this might only have a slight impact. I will fix it.

[Jiahui Zhang]

thanks for the issue, fixed.