1

I am using FCN8s and its pre-trained model for semantic segmentation on my data. Since vanilla-fcn models accept three channels images (RGB), I do not know how can I give 5 channel images to the model. Could someone please explain how can I give an input to a model that is using a pre-trained model? Thanks

S.EB
  • 1,966
  • 4
  • 29
  • 54
  • 1
    Changing the input dimension to the net will change the computed features. Unless there is a **strong** connection between the 5-channel input you are about to feed the net and the original RGB information the net was trained on, there is little sense in fine-tuning a net in this scenario. – Shai Jan 07 '18 at 06:31
  • Thanks for your comment. Since I do not have enough number of images (almost 8000) for training from scratch, the only option seems to be apply fine-tuning. Moreover, the 5 inputs are quite different from the original net. What would be your suggestion? thanks – S.EB Jan 09 '18 at 05:34
  • Might not be a complete solution to what you are looking for but you can always augment your data (take multiple crops, flip(mirroring), rotate, add noise etc.) to increase the number of samples (also diversity) and decrease overfitting – dusa Jan 10 '18 at 14:42
  • @dusa thank you very much for your help. May I ask does `caffe-master` branch provides these augmentation techniques? Could you please share some resources if you know any? Thanks once again – S.EB Jan 11 '18 at 02:31
  • @S.EB What I mentioned is partially addressed in Caffe (you can check Caffe Transformer and io.py to see how images are preprocessed). You can see this guide: https://caffe2.ai/docs/tutorial-image-pre-processing.html If you want to build on top of that, you can also add custom methods. Of course, since you have 5 channels you will have to customize how images are handled in these methods. – dusa Jan 11 '18 at 09:43

0 Answers0