How to let the Python and MATLAB wrappers work properly on classifying gray images?

[刘超]

Hi,

After checking the code and comments in these wrappers, I find that they may only run correctly when the pictures you want to classify have three channels without any modification, so I am curious if they can be changed to work well on gray pictures. I also find there is a lenet.prototext in example/mnist/, it treats pictures as its imput directly and that is different from the definition of the lenet_train_test.prototext which has to have the input data in lmdb format, so I guess there should be a way to do the classification on gray pictures as well ,but I just don't know how to do it. Here is doucument about how to use the python interface to do the classification:http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/classification.ipynb

I would be so grateful if you can tell me how to do the trick. Thanks a lot.

Happy hacking~

[Johnny ho]

You can work with grayscale image (with only 1 channel). Basically,what you do is to use /example/imagenet/create_imagenet.sh to create your lmdb / leveldm .mdb files from all your images. But in that script, you need to add --gray option, i.e., 

GLOG_logtostderr=1 $TOOLS/convert_imageset \

    --resize_height=$RESIZE_HEIGHT \

    --resize_width=$RESIZE_WIDTH \

    --shuffle \

    --gray \ 

    $VAL_DATA_ROOT \

    $DATA/val.txt \

    $EXAMPLE/ilsvrc12_val_lmdb

Then, your input can be grayscale image and it should properly.

GLOG_logtostderr=1 $TOOLS/convert_imageset \

    --resize_height=$RESIZE_HEIGHT \

    --resize_width=$RESIZE_WIDTH \

    --shuffle \

    --gray \

    $TRAIN_DATA_ROOT \

    $DATA/train_labeling_3d_axial_z_axis_length_5.txt \

    $EXAMPLE/tumor_train_lmdb_3d_axial_z_axis_length_5

[刘超]

Thanks a lot！But actually this is not my problem, maybe my description confused you, sorry about that. 

Below is a link from another caffe user and it should describe my question properly.

https://groups.google.com/forum/?utm_medium=email&utm_source=footer#!starred/caffe-users/-QT2QWoujnU

在 2014年12月13日星期六UTC+8上午4时16分59秒，Johnny ho写道：