search

DeLightCMU / CASD

85 stars 21 forks source link

Query about proposal selection for CASD #2

Open bradezard131 opened 3 years ago

bradezard131 commented 3 years ago

Hey, super interesting work. I was wondering, your paper says that N_K positive and negative proposals are selected as in OICR, and then for all of those proposals you compute your CASD attention maps/losses. OICR does no subsampling of proposals, all selective search proposals are labelled as either positive or negative and are backpropogated. Does this mean that you actually just use them all? Or do you use some other method (e.g. only the pseudo ground truths, only the positives, or using an ignore threshold as in Tang et. al. ECCV2018)?

jinhseo commented 3 years ago

Hi, I just found some trick for selecting N_K(fg/bg). FG_THRESH, BG_THRES_LO and HI determines the range of fg/bg inds. I am not sure other method at Tang et. al ECCV2018, but Kosugi et. al ICCV2019 used similar method for instance labeling.