Hello, I've found there's ratio for pos/part/landmark/neg which is 1:1:1:3 in train.py
However, in gen_hard_example.py, we'll generate neg/pos/part samples using IOU but the data numbers might be imbalanced. Ex: Pos = 80000, Part = 120000, Neg = 500000
Shall we do things like gen_imglist_pnet.py to force the ratio will work? (neg_keep = npr.choice(len(neg), size=len(neg), replace=True))
Hello, I've found there's ratio for pos/part/landmark/neg which is 1:1:1:3 in train.py
However, in gen_hard_example.py, we'll generate neg/pos/part samples using IOU but the data numbers might be imbalanced. Ex: Pos = 80000, Part = 120000, Neg = 500000
Shall we do things like gen_imglist_pnet.py to force the ratio will work? (neg_keep = npr.choice(len(neg), size=len(neg), replace=True))
Thanks