Caffe loss layer, mean and accuracy - machine-learning

Question

I have a fully convotuional network for depth estimation like this: (only upper and lower layers for the sake of simplicity):
# input: image and depth_image
layer {
name: "train-data"
type: "Data"
top: "data"
top: "silence_1"
include {
phase: TRAIN
}
transform_param {
#mean_file: "mean_train.binaryproto"
scale: 0.00390625
}
data_param {
source: "/train_lmdb"
batch_size: 4
backend: LMDB
}
}
layer {
name: "train-depth"
type: "Data"
top: "depth"
top: "silence_2"
include {
phase: TRAIN
}
transform_param {
scale: 0.00390625
}
data_param {
source: "train_depth_lmdb"
batch_size: 4
backend: LMDB
}
}
layer {
name: "val-data"
type: "Data"
top: "data"
top: "silence_1"
include {
phase: TEST
}
transform_param {
#mean_file: "mean_val.binaryproto"
scale: 0.00390625
}
data_param {
source: "val_lmdb"
batch_size: 4
backend: LMDB
}
}
layer {
name: "val-depth"
type: "Data"
top: "depth"
top: "silence_2"
include {
phase: TEST
}
transform_param {
scale: 0.00390625
}
data_param {
source: "val_depth_lmdb"
batch_size: 4
backend: LMDB
}
}
################## Silence unused labels ##################
layer {
name: "silence_layer_1"
type: "Silence"
bottom: "silence_1"
}
layer {
name: "silence_layer_2"
type: "Silence"
bottom: "silence_2"
}
....
layer {
name: "conv"
type: "Convolution"
bottom: "concat"
top: "conv"
convolution_param {
num_output: 1
kernel_size: 5
pad: 2
stride: 1
engine: CUDNN
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu"
type: "ReLU"
bottom: "conv"
top: "result"
relu_param{
negative_slope: 0.01
engine: CUDNN
}
}
# Error
layer {
name: "accuracy"
type: "Accuracy"
bottom: "result"
bottom: "depth"
top: "accuracy"
include {
phase: TEST
}
}
layer {
name: "loss"
type: "EuclideanLoss"
bottom: "result"
bottom: "depth"
top: "loss"
}
Now I have 3 questions:
When I am training the network the accuracy layer is always 1. I do not understand why?
Is EuclideanLayer the correct layer for this purpose?
Is the mean needed in such a case or can I neglect the mean?
#Define image transformers
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_mean('data', mean_array)
transformer.set_transpose('data', (2,0,1))
image = "test.png"
img = caffe.io.load_image(image, False)
img = caffe.io.resize_image( img, (IMAGE_WIDTH, IMAGE_HEIGHT))
net.blobs['data'].data[...] = transformer.preprocess('data', img)
pred = net.forward()
output_blob = pred['result']

Answer 1

Accuracy is always 1 - see this answer.
"EuclideanLoss" layer is a good fit for regression.
Subtracting the mean should help the net converge better. Keep using it. You can read more about the importance of data normalization and what can be done in that respect here.