Training Faster RCNN

27 views

Skip to first unread message

NM

unread,

Feb 24, 2017, 2:19:35 PM2/24/17

to Caffe Users

Hi Everyone,

I am trying to train Faster RCNN using VGG-16 on pascal VOC essentially trying to reproduce the results of the paper, using the python fork. I have tried both alternate training and joint approximate training method with the hyperparameters(LR =0.001) as provided in the scripts.

I however get my training loss to oscillate between 1 and 0.2 after 40,000 iterations, and the mAP on the Test set is 9.7%

I am not sure what i am doing wrong.

Help would be appreciated if anyone has reproduced the numbers or achieved similar numbers