Can you release the training code? about region-ensemble-network HOT 18 CLOSED

Hi, sorry to bother again. I am a newcomer for caffe and I have a few questions. First, When training the data, did you only add the new data and loss layer for the project to original caffe, anything else? Second, did you put the data augmentation in the data layer you edited by yourself? Is there any way else can achieve the same goal in caffe?

from region-ensemble-network.

@qhykwsw Hi, sorry for the late response.

First, When training the data, did you only add the new data and loss layer for the project to original caffe, anything else?

We added a new data layer, a loss layer and add more augmentation functions in the data transformer.

Second, did you put the data augmentation in the data layer you edited by yourself?

We put them into the data transformer.

from region-ensemble-network.

I see, thanks for your advices. Can you share the new data layer file to me? My email address is [email protected]. If you cannot, does it need to write .cu file for the new data layer?

from region-ensemble-network.

@qhykwsw Hi, the data layer contains lots of other things that are used in my lab, so I'm afraid that I could not share it with you now. Sorry for that. It is only for CPU, not .cu.

from region-ensemble-network.

Nevermind, I guess I can take this opportunity to learn caffe. Thanks again and wish you do better research.

from region-ensemble-network.

@qhykwsw Thanks. I think the results are not difficult to reproduce. Welcome to ask any questions if necessary.

from region-ensemble-network.

Thanks for your patience. In fact, I do have two question. First, I am curious about when do the data augmentation? Is that after the cube was cropped or directly on the orign picture? Second, if you do the data augmentation after the cube was cropped, when doing the translation within [−10, 10] pixel, the joints may become out of the cube. What should I do about this situation?

from region-ensemble-network.

First, we did it after cropping to reduce the times of transformation.

Second, I think it didn't harm the network too much when the joints are slightly out of the cube. And you can also set the cube large enough to avoid this situation.

from region-ensemble-network.

Ok, I see. So, what's the best augmentation order? First translation and then scaling and then rotation? Or it's really doesn't matter? And when doing the scaling and rotation, what's the reference point, center of the picture, or the center of mass calculated by the previous step?

from region-ensemble-network.

It doesn't matter. You can try it:) The center is the one of cropping.

from region-ensemble-network.

Ok, last question. When doing the augmentation, for example translation, there will be some pixels newly go into the cube, how should I set these pixels, 0, 255 or any value else?

from region-ensemble-network.

It depends. We set them to background values.

from region-ensemble-network.

I see and I will try to reproduce the research follow your setting, thank you so much.

from region-ensemble-network.

Hi, I am also trying to augment my data. Just curious, after translating the depth image, eg to the right by 10 pixels, do you also change the ground truth joint (x,y,z) coordinate to reflect that?

if yes, how do you do it? do you just add +10?

from region-ensemble-network.

@dhecloud Sure, the joint should also be changed by add 10. Otherwise the results are wrong.

from region-ensemble-network.

Hi, I noticed that the dropout layer in your model(deploy.prototxt) didn't specify the phase. Does that mean the model do dropout in both training and testing phase?

from region-ensemble-network.

@qhykwsw The layer itself will take care of it. You can view the layer codes for details.

from region-ensemble-network.

I did not expect this feature. I specify the dropout layer just work in the trainning phase as usual.

from region-ensemble-network.