DirtyHarryLYL / DJ-RN

As a part of HAKE project (HAKE-3D). Code for our CVPR2020 paper "Detailed 2D-3D Joint Representation for Human-Object Interaction".
Apache License 2.0
100 stars 13 forks source link

How can I get the GT and Neg data? #41

Closed Ivan-VV closed 3 years ago

Ivan-VV commented 3 years ago

Would you like to explain about the GT and Neg data in step5 of Data generation. In the previous steps, I only get the dataset divided into train and test. Thank you very much!

DirtyHarryLYL commented 3 years ago

Here, GT and neg indicate the positive/interactive and negative/non-interactive human-object pairs. GT pairs are recorded in the HICO-DET labels. Negative pairs can be paired automatically, e.g., in an image, we can exhaustively pairing all detected humans and objects, then the pairs except for the GT pairs are the neg pairs.

Ivan-VV commented 3 years ago

Thanks for your reply! So are GT and Neg samples used to learn interactiveness knowledge mentioned in your TIN paper? Is it right that the GT and Neg samples are assigned automatically when I run the scripts in step5 of Data generation?

DirtyHarryLYL commented 3 years ago

Yes, GT and Neg pairs can be leveraged to learn the interactiveness and HOI classification. In the HOI classification, normally the pos:neg sample ratio is about 1:4. The GT and Neg pairs with the corresponding human poses are preprepared in our file, so you could directly use them except that you want to reorganize the neg pairs for your convenience.

Ivan-VV commented 3 years ago

Thx!