demo code

[buaaswf]

Dear @meliketoy,
We attempt to draw the interaction results in the images, could you please tell us that whether the inst_target is the interaction actions?

[bmsookim]

@buaaswf The inst_target is a target that is used to evaluate the AP_agent metric which is no longer used in HOI detection. The interaction labels will be in the name of target['pair_action'].

[buaaswf]

Dear @meliketoy,
Thanks for your kind help, could you please tell me more about the HOI detection demo process?
I am now just trying to use the pretrained model to predict the HOI labels of a single image.

[buaaswf]

Dear @meliketoy,
I find that the VCOCO's action labels have 29 classes, while the pair actions only have 26 classes, could you tell me to remove which 3 classes?

[bmsookim]

@buaaswf The 4 classes that are excluded are the classes that do not include a target object (intransitive verbs), which are 'walk', 'run', 'smile', 'stand', and 1 class is additionally excluded during the evaluation due to its very small sample number, 'point'. Interactions including 2 target objects are treated as separate interactions.
Therefore, there are a total of 25 actions, and the final verb is an additional target value for 'no interaction', which is explained in detail in our paper :)

Refer to our paper and VSRL (https://arxiv.org/pdf/1505.04474.pdf) for further information about the dataset.