search

YueLiao / CDN

Code for "Mining the Benefits of Two-stage and One-stage HOI Detection"
Apache License 2.0
89 stars 15 forks source link

Ablation experiments for CDN #27

Closed hutuo1213 closed 1 year ago

hutuo1213 commented 1 year ago

Hi, Dr.Liao,

Thank you for your excellent work of your CDN. We have a problem where you report the performance of different variants on the test set in the CDN to choose the best framework. For example: 1 Do you divide the validation set for Hico-Det, and if not, can you tell me why you did it? We have encountered unprecedented difficulties.

Thank you very much for your consideration. I look forward to hearing from you soon.

Sincerely yours, yaoyaosanqi.

YueLiao commented 1 year ago

Hi,

This answer can be found in the Openreview Reviewer LC3p, Q12.

hutuo1213 commented 1 year ago

Sorry I didn't make it clear. I wonder if it is possible to tune hyperparameters based on test set on hico-det?

YueLiao commented 1 year ago

We have not tuned hyperparameters based on the test set on hico-det while just showing ablation studies on the HICO-DET test set. However, we think it is somewhat reasonable to tune the hyperparameters based on the HICO-DET test set while its test set plays a validation set role in most HOI detection research papers.

hutuo1213 commented 1 year ago

Thank you for your reply.

hutuo1213 commented 1 year ago

We can't tell the essential difference between the two, please let me know. The modules we are proposing are valid, however, I would like to know how many blocks are added consecutively to achieve the best performance. How do I do this?

YueLiao commented 1 year ago

The best and most reasonable way is to divide a minimum validation set from the training set to validate the effectiveness of different modules. However, many HOI detection papers directly conduct such experiments in the test set.

hutuo1213 commented 1 year ago

Thanks a lot