How can i specify grayscale image in caffe?

[蒋楠]

Hi,

I am using caffe with the grayscale image as input. I use the tool: convert_imageset to create the lmdb dataformat.

However, when I use my own CNN netowrk, with the data layer as following, the channel size of the first blob is 3, just like I am using the 3 channel image. The grayscale image is synthesized by myself in matlab, and when i load it in the matlab, it is also single channel. But why the caffe specify it as 3 channel? does it because I wrongly synthesized the grayscale image? or because i wrongly used the tool: convert_imageset ?

Has anybody met this problem before?  I appreciate all the reply you provide.

layer {

  name: "image"

  type: "Data"

  top: "dXataX"

  top: "ylabely"

  include {

    phase: TEST

  }

  transform_param {

    scale: 0.004                     #change here

  }

  data_param {

    source:  "../../data/syn_reshaped/train_lmdb"

    batch_size: 256

    backend: LMDB

  }

}

layer {

  name: "image"

  type: "Data"

  top: "dXataX"

  top: "ylabely"

  include {

    phase: TRAIN

  }

  transform_param {

    scale: 0.004                    #change here

  }

  data_param {

    source: "../../data/syn_reshaped/cv_lmdb"

    batch_size: 256

    backend: LMDB

  }

}

[Venu Sharma]

Hello Nina,
   I think once you set the bool variable gray ON in convert_imageset and create LMDB your issue will be solved.

Regards

[Tarun Sharma]

While calling convert_imageset, do convert_imageset --gray [training data location] [location to train.txt] [location where lmdb will be created]

On Friday, July 3, 2015 at 7:40:36 AM UTC+5:30, Nina wrote:

[Oscar Beijbom]

you could also consider using a python data-layer for full control.

http://nbviewer.jupyter.org/github/BVLC/caffe/blob/master/examples/pascal-multilabel-with-datalayer.ipynb

Oscar Beijbom

[Chi Zhang]

Thanks Tarun. I found that with adding "--gray" the LMDB was created as single-channel but the values in each grayscale image are mormalized somehow. For example, in PASCAL VOC 2012, the segmentation label are PNG files with values 0, 1~20, 255, but I got some value like 90, 108, 113... Is there a way to keep original values as what it was in PNG images?

Thanks in advance.

Best,

Chi Zhang