Deformable-ConvNets
Deformable-ConvNets copied to clipboard
detector output is not that consistant
Hi
I run FPN version of it for testing images provided in /demo.
I observed that detector boxes output changes for every different run without changing script or config.
And also observed that getting many ROI's for single object as shown in figure below. And GPU memory keeps on increasing for every image.
Do I have to do any modifications in the NMS threshold to get consistent and single ROI for objects.
