How to train the model by myself ?

[thesby]

I want to use the crfasrnn model to train on my data, but there is no network for training. Should I just add a data layer to provide label, and add a softmax layer to get loss ? If not, please give some suggestions about it. Thank you.

[bittnt]

https://github.com/bittnt/caffe/tree/22babdb4a70eb5829c52fccc417d4b2f916d2748/examples/segmentationcrfasrnn

[westeast]

@thesby Have you trained your model by using your data?

[thesby]

@westeast Yes, I have trained one.

[chitrang6]

Hey I have a data set of cars and non cars image and using that I want to detect cars on the highway. How to train using my dataset? Kindly help

On Nov 28, 2016 11:29 PM, "Luo_yuansheng" [email protected] wrote:

@westeast https://github.com/westeast Yes, I have trained one.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/torrvision/crfasrnn/issues/48#issuecomment-263496728, or mute the thread https://github.com/notifications/unsubscribe-auth/AK0P9L4ir4Kx4sOxE4mC5gYvzskY-8Ujks5rC9ROgaJpZM4IXDPz .

[thesby]

So you want to do object detection or image segmentation ? If you just want to segment images, just use image data layer to provide images, and add a crf layer as the paper did.

[marticsvk]

@thesby Hi thesby, i'm currently working on training my own data set based on this CNN-CRF model for image segmentation. Could you please write me your email address so we can communicate there or directly send me a message at [email protected], i'd appreciate your help.

[matthiasmace]

@bittnt Hi bittnt ! I want to train on my data (animal art representation). I followed the link mentionned above. It is pretty easy to understand but could you please explain the following in details :

rgb-train" label-train" rgb-val" label-val"

Many thanks ! By the way, it now works fine on my Macos 10.12 !

[bittnt]

These are the folder names of lmdbs. rgb-train is the lmdb folder for the input image data, while the label-train is for the label map.

You should be able to create lmdb following the caffe fashion.

[matthiasmace]

@bittnt OK, thanks. But What do I need to put inside label-train please ? Do I follow the indications given in this link ? http://host.robots.ox.ac.uk/pascal/VOC/voc2012/segexamples/index.html And if yes, "object" or "class" ? Thanks again

[bittnt]

class.

object is for instance segmentation.