Regarding the Negative Samples

[yunfan0621]

Hi Weilun Wang,

Thank you for sharing the code of your great work!

I am wondering is it possible to visualize the 'negative samples' by re-projecting them into the image space? Or they only virtually exist in the latent space to help the encoder learn more discriminative features? I was expecting to see some visualization of the generated hard negative examples while reading the paper, but failed to do so.

Best,
Yunfan

[WeilunWang]

Hi Yunfan,

Thanks for your interest.

Since it is hard to reverse the encoding process, we cannot directly decode the generated negative examples.
We visualize of negative examples by retrieving regions based on generated features in Figure 5.
More precisely, we calculate the distance between the generated negative sample and examples of all regions, and use the closest region as the visualization of the negative sample.

Best,
Weilun