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
32 stars 3 forks source link

Trouble producing the interactiveness AP & HOI mAP reported in the paper. #4

Closed Charles-Xie closed 1 year ago

Charles-Xie commented 1 year ago

Thanks for your awesome work! I try to reproduce the results reported in your paper recently on hico-det dataset, but meet 2 problems:

  1. Using the config & training command, I cannot produce the interactiveness AP in 1st training stage. I follow exactly the steps in your readme, during 100 epochs of training & validation, the interactiveness AP is 37.47 at most. The interactiveness AP reported in the paper is 38.74. (The training & validation log is here log.txt.) What could be the reason?

  2. The training config & command for 2nd training stage is not provided. According to your paper, the training consists of 2 stage, first with det & interactiveness loss and second with det & verb loss. The provided training config & command is for the 1st stage (interactiveness cls), and that for the 2nd stage is missing, so the verb cls results and final HOI mAP cannot be reproduced.

I look forward to your reply! Thank you!!! 🙏

enlighten0707 commented 1 year ago

Hello! We have provided the prediction results of 1-st and 2-nd stage, and auxiliary benchmark code to output final HOI mAP.