This feature author did not write tutorial, so I wrote, not to repeat, but to supplement, author of the tutorial official website, must look, must take notes to write down what to change:
Http://mi.eng.cam.ac.uk/projects/segnet/tutorial.html
First need to install Caffe, and can make runtest success, no error https://github.com/alexgkendall/caffe-segnet, Note that if you want to use cudnn5.1 acceleration, you should use the URL https://github.com/TimoSaemann/caffe-segnet-cudnn5 provided by the author. This caffe refers specifically to the author's change, not under the official website. after Download the network code https://github.com/alexgkendall/SegNet-Tutorial provided by the author.
Because we want to use the training network parameters provided by the author, or weight, so we need to download this: https://github.com/alexgkendall/SegNet-Tutorial/blob/master/Example_Models /segnet_model_zoo.md
It is recommended not to ignore the author's readme in English and any English.
Then all formats are Caffemodel and can be used.
What are you going to change then, of course 1 training statements: What prototxt you want to use, 22. py files, and all 3 paths used. In other words, to insure, you can open all the prototxt and change the absolute path to a relative path. Also note that your 4 file level is not the same as the author, not the same relative path to change. At the same time, if it is a 5cpu run code also note that the GPU in. Py is also changed to the GPU in the Cpu,6.prototxt file to be CPU.
Again about the function if you want to change the output form by 7 you can change Test_segmentation_ Camvid.py inside the code, such as you want to output their segment results you put scipy that line of comments, and then change the parameter RGB to label, the name is also written as ' media\tinae\+ ' I ' + ' _segnet ' similar to this naming method.
All right, finish.
Test your diagram with the Segnet author's trained network