Question1 about loss define.

[taoruijie]

Question1 about loss define.

[TaoRuijie]

Q1: We follow the original AVA paper, which choose 0.4.Check here:https://arxiv.org/pdf/1901.01342.pdf, Page 7. Similar setting can be found in this paper: https://openaccess.thecvf.com/content_CVPR_2020/papers/Alcazar_Active_Speakers_in_Context_CVPR_2020_paper.pdf, page 5, they choose 1.0.

Q2: As we mentioned in Q1, in many existing works already used 3 losses togethor. Adding this into the figure will make it difficult to understand, so we ignore that.

From my experments, this setting can improve the performance by a small margin (about 0.5% mAP)

[TaoRuijie]

Also for prediction, in our work we use the outsAV only. Three losses are only used for training as what I remember.