ppengtang / pcl.pytorch

PyTorch codes for our papers "Multiple Instance Detection Network with Online Instance Classifier Refinement" and "PCL: Proposal Cluster Learning for Weakly Supervised Object Detection".
MIT License
253 stars 48 forks source link

About result of vgg16_voc2007.xml #51

Closed Name4H closed 3 years ago

Name4H commented 3 years ago

The result of vgg16_voc2007.xml with WITH_FRCNN:Fasle map is 50.06.and WITH_FRCNN:True map is 51..50,Is this normal?

ppengtang commented 3 years ago

Hi, sorry I don't remember the numbers of WITH_FRCNN: False. For WITH_FRCNN: True, mAP 51.50 should be reasonable. You could try different random seeds (1~10) to reproduce the reported number.

Name4H commented 3 years ago

thanks,I will try

hhhh12345678 commented 3 years ago

Hi, I am getting troubles with the result of vgg16_voc2007.

I used the default hyper-parameters to train the net, but got only 45.48mAP with WITH_FRCNN on.

I have read relate issues about performance dropping, but I found the results I got seems much lower.

my environment as below:

cuda 10.1 + pytorch 1.5.1 + RTX2080Ti