Skip to content

Latest commit

 

History

History
98 lines (65 loc) · 1.7 KB

README.md

File metadata and controls

98 lines (65 loc) · 1.7 KB
Raw

picture-clustering

This source code obtains the feature vectors from images and write them in result.csv.

After that you cluster feature vectors by unsupervised clustering (as clustering_example.py).

unsupervised clustering example: SpectralClustering, k-medoids, etc ...

notice

you need meanfile, modelfile, and networkfile.

I recommend to use Imagenet model in Caffe.

  • meanfile: ilsvrc_2012_mean.npy
  • modelfile: caffe_reference_imagenet_model
  • networkfile: imagenet_deploy.prototxt.

Please, rewrite there path(line 16 ~ 18).

How to get there files.

  • caffe_reference_imagenet_model
wget https://raw.githubusercontent.com/sguada/caffe-public/master/models/get_caffe_reference_imagenet_model.sh
chmod u+x get_caffe_reference_imagenet_model.sh
./get_caffe_reference_imagenet_model.sh
  • imagenet_deploy.prototxt
wget https://raw.githubusercontent.com/aybassiouny/wincaffe-cmake/master/examples/imagenet/imagenet_deploy.prototxt

  • ilsvrc_2012_mean

    ilsvrc_2012_mean.npy is in caffe/python/caffe/imagenet/ilsvrc_2012_mean.npy

rewrite imagenet_deploy.prototxt

line 174: "fc6" -> "fc6wi"  
line 186: "fc6" -> "fc6wi"

It shows an example.

   
layers {
  name: "fc6"
  type: INNER_PRODUCT
  bottom: "pool5"
  top: "fc6wi"
  blobs_lr: 1
  blobs_lr: 2
  weight_decay: 1
  weight_decay: 0
  inner_product_param {
    num_output: 4096
  }
}
layers {
  name: "relu6"
  type: RELU
  bottom: "fc6wi"
  top: "fc6"
}
layers {

files

input

put the image in the folder './data'.

  • ./data/picture_images

output

  • ./result.csv

if you do clustering_example.py,

  • ./result

usage

python feature.py

and you want to try clustering example.

python clustering_example.py