image-level中scale1和scale2是怎么来的 about hake-action HOT 2 CLOSED

Do you mean the 117 and 80 classification scores from two separated classifiers?
These two classifiers are additional tasks to help the 600 classification, we found they can boost the performance sometimes. And after transferring the 117 or 80 to 600 scores via the prior relation matrix from Hico-det, the performance would be further boosted.
The fusing factors can be obtained via grid search on Val set, another way is to build a dynamic way to learn the factors, i.e., use the verb and object features -->MLP-Softmax-->factors, and then operate fusion. These two ways work depend on different scenarios, you may need trail-and-error to find the best way.

from hake-action.

OH. I see. The dynamic way seems cool. Thanks a lot.

from hake-action.