caffe pixel-wise classification / regression

问题

What I want to do is to do a simple pixel-wise classification or regression task. Therefore I have an input image and a ground_truth. What I want to do is to do an easy segmentation task where I have a circle and a rectangle. And I want to train, where the circle or where the rectangle is. That means I have an ground_truth images which has value "1" at all the locations where the circle is and value "2" at all the locations where the rectangle is. Then I have my images and ground_truth images as input in form of .png images.

Then I think I can either to a regression or classification task depending on my loss layer: I have been using the fully convolutional AlexNet from fcn alexnet

classification:

layer {
  name: "upscore"
  type: "Deconvolution"
  bottom: "score_fr"
  top: "upscore"
  param {
    lr_mult: 0
  }
  convolution_param {
    num_output: 3 ## <<---- 0 = backgrund 1 = circle 2 = rectangle
    bias_term: false
    kernel_size: 63
    stride: 32
  }
}
layer {
  name: "score"
  type: "Crop"
  bottom: "upscore"
  bottom: "data"
  top: "score"
  crop_param {
    axis: 2
    offset: 18
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss" ## <<----
  bottom: "score"
  bottom: "ground_truth"
  top: "loss"
  loss_param {
    ignore_label: 0
  }
}

regression:

layer {
  name: "upscore"
  type: "Deconvolution"
  bottom: "score_fr"
  top: "upscore"
  param {
    lr_mult: 0
  }
  convolution_param {
    num_output: 1 ## <<---- 1 x height x width 
    bias_term: false
    kernel_size: 63
    stride: 32
  }
}
layer {
  name: "score"
  type: "Crop"
  bottom: "upscore"
  bottom: "data"
  top: "score"
  crop_param {
    axis: 2
    offset: 18
  }
}
layer {
  name: "loss"
  type: "EuclideanLoss" ## <<----
  bottom: "score"
  bottom: "ground_truth"
  top: "loss"
}

However, this produces not even the results I want to have. I think there is something wrong with my understanding of pixel-wise classification / regression. Could you tell me where my mistake is?

EDIT 1

For regression the retrieval of the output would look like this:

output_blob = pred['result'].data

predicated_image_array = np.array(output_blob)
predicated_image_array = predicated_image_array.squeeze()
print predicated_image_array.shape
#print predicated_image_array.shape

#print mean_array
range_value = np.ptp(predicated_image_array)
min_value = predicated_image_array.min()
max_value = predicated_image_array.max()

# make positive
predicated_image_array[:] -= min_value

if not range_value == 0:
    predicated_image_array /= range_value

predicated_image_array *= 255
predicated_image_array = predicated_image_array.astype(np.int64)
print predicated_image_array.shape

cv2.imwrite('predicted_output.jpg', predicated_image_array)

This is easy since the output is 1 x height x width and the values are the actual output values. But how would one retrieve the output for classification / SotMaxLayer since the output is 3 (num labels) x height x width. But I do not know the meaning of the content of this shape.

回答1:

first of all, your problem is not regression, but classification!

if you want to teach the net recognise circles and rectangles you have to make a different data set - an images and labels, for example: circle - 0 and rectangle - 1. you do it by making text file that containsthe images path and the images labels, for example: /path/circle1.png 0 /path/circle2.png 0 /path/rectangle1.png 1 /path/rectangle1.png 1. here is a nice tutorial for a problem like yours. good luck.

来源：https://stackoverflow.com/questions/40549334/caffe-pixel-wise-classification-regression