How about performance evaluation result

[John1231983]

Thanks for sharing your work!. I have some information that may be useful to you.

1. You are using SGD. Any reason why did not use Adam method? You can have better performance by using Adam with the following setting

   lr_policy: "fixed"
   gamma: 0.1
   base_lr: 0.0001 
   momentum: 0.99
   # no gradient accumulation
   iter_size: 1
   max_iter: 100000
   weight_decay: 0.0005
   type:"Adam"

2. For infant image segmentation, do you use original intensity or normalization intensity? Which kind of normalization method did you use? For example, (1) normalize in zero mean and unit variance, (2) normalize in the range [0,1] by divide the image to maximum intensity (1000) .

3. I am interested in your performance in the testing set of infant dataset (from subject 11 to subject 23). Could you give me how much performance did you achieve?

[ginobilinie]

@John1231983 Thank you.

1. Thanks for your share. I also use Adam later in this project.

2. normalized intensity: (x-mu)/(max-min)

3. what do you mean by sub11 to sub23? which dataset?

Thanks.

[John1231983]

Thanks for your information. I am regarding to the infant segmentation that used in the iseg dataset (http://iseg2017.web.unc.edu/). Did you try it in your paper? The testing set includes from subject 11 to subject 23

[ginobilinie]

@John1231983 This is a new dataset which differs from the datasets I use in the paper. I have done sth with the dataset, I can check it for you a little later.

[John1231983]

I see. I was misunderstood between infant and iseg dataset. For second question, why not use normalize with zero mean and unit variance which are so popular with using BatchNorm. It means im=(im-im.mean())/im.std()

[ginobilinie]

@John1231983 Thanks for your suggestion.