Problems in using other backbone networks

Hi, I tried to replace the VGG backbone of mscnn with resnet 50, but when I train the 7s-576-2x model on kitti car dataset, it gets quite bad performance:

only 56% box recall on the output of stage1(iou threshold 0.7), which is around 97-98% for the vgg conterpart and the log shows that the networks fits well on the kitti car . It really confused me much.

I kept all the other parameters same as vgg, such as feature map size of every branch and anchor size. Have you tried using resnet or other backbone network ?

zhaoweicai / mscnn

Problems in using other backbone networks #80