search

ashafaei / dense-depth-body-parts

Dense classification of the depth images to recognize the body parts.
http://www.cs.ubc.ca/~shafaei/homepage/projects/crv16.php
GNU General Public License v3.0
42 stars 21 forks source link

Number of output of deconvolution or convolution layer #5

Closed gizmo1904 closed 7 years ago

gizmo1904 commented 7 years ago

Dear Mr. Shafaei, I was looking into your model and I am confused by the number of outputs in different layers. In the deploy.prototxt the output size is 46, in the model overview it's 44 (43 classes + background). Could you please explain that to me? I've got your model from here: http://www.cs.ubc.ca/~shafaei/homepage/projects/datasets/caffe_crv16.zip

Thank you very much!

ashafaei commented 7 years ago

Hello @gizmo1904,

Thank you for your question.

The texture that I used to generate the data has 45 regions (see the dataset UBC3v and the 3d models and textures 3dmodels.tar.gz). However, two of the regions in the texture file never shows up in the rendered images. When I started the training, I was not aware that the two classes in the texture are actually never rendered, which lead to two useless classes in the output of the network. I later simply ignored the two classes as they never became activated anyway. You can easily identify the extra useless classes by examining the activations.

Let me know if I could provide more help.

Cheers, Alireza

gizmo1904 commented 7 years ago

Thank you for clearing that up. Have a nice day.