Skip to first unread message
César Salgado
Nov 14, 2015, 6:44:57 AM11/14/15
to Caffe Users
I want to use a trained convnet to predict labels for every pixel of an image. I have already seen this notebook example: http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/net_surgery.ipynb
The problem with this notebook example is that it doesn't generate an output for every window. It just gives an 8x8 prediction map instead of 451x451 prediction map.
I am also aware of FCN in the model zoo: https://github.com/longjon/caffe/tree/future
But I was looking for a simpler solution than to try to figure out how to use that branch. Besides, I don't code to train. I just want to do inference.
Is there a way to run a trained convnet densely in an image in a way faster than the naive approach, ideally just using the python interface and just using the caffe master branch?
The problem with this notebook example is that it doesn't generate an output for every window. It just gives an 8x8 prediction map instead of 451x451 prediction map.
I am also aware of FCN in the model zoo: https://github.com/longjon/caffe/tree/future
But I was looking for a simpler solution than to try to figure out how to use that branch. Besides, I don't code to train. I just want to do inference.
Is there a way to run a trained convnet densely in an image in a way faster than the naive approach, ideally just using the python interface and just using the caffe master branch?
Reply all
Reply to author
Forward
0 new messages