Question about the Trainval_Neg_HICO.pkl about transferable-interactiveness-network HOT 5 CLOSED

We do not generate this training file by ourselves, please refer to the iCAN website.

from transferable-interactiveness-network.

Thank you for your reply!
I also have a question that how to calculate the loss for negative samples, for the wrong relationship between the human and object?

from transferable-interactiveness-network.

Positive samples and negative samples differ in the ground truth label. Loss computation is the same (softmax/cross_entropy) between these two.

from transferable-interactiveness-network.

Negative samples have wrong relationships. If loss computation is same as positive samples, minimizing loss does not mean the predictions for these negative samples tend to the wrong relations?

from transferable-interactiveness-network.

Exactly speaking, HOI detection is a multi-label classification problem. 600 HOI classes are separate during the inference (600 sigmoids). So a pair can be pos to one class, and neg to another class. In our paper, we focus on the interactive or non-interactive pair problem. That is to say, whether a pair has one or more HOIs or zero HOI.

from transferable-interactiveness-network.