rbgirshick/rcnn

lower performance than expected

Opened this issue · 3 comments

Hi,

I have used the RCNN_model that you have provided to get the results on VOC 2007. The thing is I could get MAP = 42.71 with the provided model, while in the paper the lowest claim is 44.2. Also, you claimed MAP = 66 for RCNN in Fast RCNN paper while I have got 40.11 when I trained my network. Since I am using your code without any modification, I think the problem is with the networks that I have used. Is it possible for you to release the trained rcnn models (mat files) that you have used for VOC 2007, 2010 and 2012 using Alexnet and VGG16?

Thanks so much.

@eefami Have you used the code on Windows?

@Addhi86 No. I used it on Linux Ubuntu.

I am using Windows and I am having a problem to run the code.
I have sent you an email regarding the Object detection on research gate. Can you, please reply me there