I was having a problem for configuring caffe to do pixel wise segmentation. However, after research, I found that you can create another Imdb database for labels. I have done that, but I don't know how to link the data and label databases in the training prototxt file. Any example would be much appreciated.
Asked
Active
Viewed 288 times
0
zx485
- 28,498
- 28
- 50
- 59
Ahmed Nadeem
- 33
- 4
-
Included links of referred providers and fixed some grammar. – zx485 Aug 08 '16 at 13:51
1 Answers
0
Since you are aiming for pixel-wise segmentation I'm gonna call what you call labels as ground truths.
The simple solution is - Use two input layers.
Use one input layer to read the image and some fake labels. Use another input layer to read the ground truth and some fake labels. Use the image and ground truth blobs wherever appropriate in your network.
malreddysid
- 1,342
- 1
- 9
- 12
-
how will i configure that in the net.prototxt file?, i can't find an example of that – Ahmed Nadeem Aug 08 '16 at 13:30
-
Take the alexnet prototxt file for example. Instead of one input layer. Put two. – malreddysid Aug 08 '16 at 13:38