What is the reason for using negative examples?

[mahsa1363]

Hello
Negative samples have wrong relationships. Therefore minimizing loss does not mean the predictions for these negative samples tend to the wrong relations?What is the reason for using negative examples?

[Foruck]

Negative samples are not 'wrong' samples. Instead, they are samples that belong to the 'no interaction' categories. Since the original dataset provides a limited annotation for 'no interaction' samples, with these additional negative samples, we can boost our model's ability of distinguishing the difference between samples with interaction and without interaction, in other words, with 'no interaction'.

[mahsa1363]

thank you for your answer
Tاhis means, all HOI classes in negative samples are 'no interaction' which related to specified object?

[Foruck]

Yes, that is correct.

[mahsa1363]

thank you very much