search

enlighten0707 / Body-Part-Map-for-Interactiveness

Code for ECCV2022 Paper "Mining Cross-Person Cues for Body-Part Interactiveness Learning in HOI Detection"
MIT License
33 stars 3 forks source link

Verb scores are always the same #8

Closed yuzheyao22 closed 9 months ago

yuzheyao22 commented 1 year ago

If I print the "vs"(verb_scores) in hoi.py, logits for all the 117 verbs are the same. It means that the model does not predict the verb, right?

yuzheyao22 commented 1 year ago

If I print the "vs"(verb_scores) in hoi.py, logits for all the 117 verbs are the same. It means that the model does not predict the verb, right?

For example when printing the _outputs['pred_verblogits'][0][0] tensor([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], device='cuda:2')

enlighten0707 commented 9 months ago

Yes, the provided code output interactiveness predictions, which are used to apply non-interaction suppression (NIS) on verb classification results. Here "pred_verb_logits" is omitted.