dual supervision for sparse/dense images

[SherlockHolmes221]

why need dual supervision for sparse/dense images

In loss fuctuin dual supervision for sparse/dense images loss_ce_sp = (raw_ce_loss weights div_mask)[sp_idx].sum() / ((weights div_mask)[sp_idx].sum() + eps) loss_ce_ds = (raw_ce_loss weights div_mask)[ds_idx].sum() / ((weights div_mask)[ds_idx].sum() + eps)

thanks

[cxliu0]

As explained in the paper (Page 6), we use dual supervision to prevent the loss from being diluted by samples with few people, because such samples yield small training losses.

[SherlockHolmes221]

thanks