Why IOU balanced sampling reduces accuracy instead？

Open ModestYjx opened this issue 4 years ago • 1 comments

I add IOU balanced sampling module on BFP_BLL strategy, but m-ap is reduced.
I want to know why this result happend. Here are my experiment results:

experiment_BFP_BLL

([('bbox_mAP', 0.37), ('bbox_mAP_50', 0.578), ('bbox_mAP_75', 0.402), ('bbox_mAP_s', 0.221), ('bbox_mAP_m', 0.405), ('bbox_mAP_l', 0.465), ('bbox_mAP_copypaste', '0.370 0.578 0.402 0.221 0.405 0.465'), ('mAP', 0.425), ('mAP_50', 0.682), ('mAP_75', 0.455), ('mAP_s', 0.281), ('mAP_m', 0.484), ('mAP_l', 0.595)])

experiment_BFP_BLL_IOUBS

([('bbox_mAP', 0.364), ('bbox_mAP_50', 0.567), ('bbox_mAP_75', 0.395), ('bbox_mAP_s', 0.209), ('bbox_mAP_m', 0.399), ('bbox_mAP_l', 0.462), ('bbox_mAP_copypaste', '0.364 0.567 0.395 0.209 0.399 0.462'), ('mAP', 0.422), ('mAP_50', 0.673), ('mAP_75', 0.451), ('mAP_s', 0.271), ('mAP_m', 0.477), ('mAP_l', 0.596)])

If you need my code, I could send it to you to check this problem .

Mar 31 '21 03:03 ModestYjx

I add IOU balanced sampling module on BFP_BLL strategy, but m-ap is reduced.

I want to know why this result happend. Here are my experiment results:

experiment_BFP_BLL

([('bbox_mAP', 0.37), ('bbox_mAP_50', 0.578), ('bbox_mAP_75', 0.402), ('bbox_mAP_s', 0.221), ('bbox_mAP_m', 0.405), ('bbox_mAP_l', 0.465), ('bbox_mAP_copypaste', '0.370 0.578 0.402 0.221 0.405 0.465'), ('mAP', 0.425), ('mAP_50', 0.682), ('mAP_75', 0.455), ('mAP_s', 0.281), ('mAP_m', 0.484), ('mAP_l', 0.595)])

experiment_BFP_BLL_IOUBS

([('bbox_mAP', 0.364), ('bbox_mAP_50', 0.567), ('bbox_mAP_75', 0.395), ('bbox_mAP_s', 0.209), ('bbox_mAP_m', 0.399), ('bbox_mAP_l', 0.462), ('bbox_mAP_copypaste', '0.364 0.567 0.395 0.209 0.399 0.462'), ('mAP', 0.422), ('mAP_50', 0.673), ('mAP_75', 0.451), ('mAP_s', 0.271), ('mAP_m', 0.477), ('mAP_l', 0.596)])

If you need my code, I could send it to you to check this problem .

i have this problem too why it makes lower ap???

Sep 08 '21 11:09 mmoghadam11