How to select correct HOI from multiple predictions during test time. about transferable-interactiveness-network HOT 1 CLOSED

Hi,
in fact, the ''right" prediction is decided by the thresholds or rank. And different metrics may have different rules.
For example, for the object apple, action hold, I can set the threshold as 0.85.
Then all the predictions with action score higher than 0.85 and the box IoU higher than 0.5 (h and o both) are "right".
Or I can also rank all the prediction of HOI "hold apple" of all testing images and choose the top-k predictions as the "right" predictions (boxes should be accurate too).
The detailed metric about how cal the mAP could be found in the evaluation files, they may help you too.
Or you can also only visualize the "true positive" found by the evaluation code as the "right" predictions.

from transferable-interactiveness-network.