questions about mAP results of anet dataset

[yangmin666]

In the paper, your results is 35.6%. However, you provide three different tools("builtin" "cuhk" "gt"), which one produce results of 35.6%

[klauscc]

Sorry for the late response. builtin should produce "35.6%".

• buildtin means using the classifier learned by the model.
• cuhk is the classifier that used by many previous work that only have binary predictions (action or not). The scores for each ANet video is in file data/annots/anet/cuhk_val_simp_share.json. I copied it from some previous work.
• gt is the results using the ground-truth classes. It should be the upper-bound of our method.

[EmreOzkose]

Hi @klauscc , if you don't mind, can you give some details about cuhk_val_simp_share.json. I don't understand what it is storing.