Inquiry about your predicted person boxes for experiments

facebookresearch / video-long-term-feature-banks

Long-Term Feature Banks for Detailed Video Understanding

Apache License 2.0

373 stars 62 forks source link

Inquiry about your predicted person boxes for experiments #4

Closed HuangOwen closed 5 years ago

HuangOwen commented 5 years ago

Thanks for your great work! I have a question of your predicted person boxes for experiments, I notice that in ava_train_predicted_boxes.csv there is action label from 1 to 80 while some of them are labeled as -1(no action perhaps?). I was wondering what is the criterion of the 'no action' human instance. Do you pick them out by IOU with GT boxes of training set? Thanks and looking forward to your reply!

dingli-dean commented 5 years ago

Maybe "-1" stands for the mismatched boxes when comparing the action labels of detected boxes with that of ground truth?

chaoyuaw commented 5 years ago

Hi, thanks for the questions!

We assign labels of a ground-truth box to a predicted box if they overlap with IoU at least 0.9. A predicted box might have no labels assigned. In that case the label is "-1".

HuangOwen commented 5 years ago

That's reasonable, thanks for your reply!