normalize_layer about refinedet HOT 6 CLOSED

@auroua , Hi
The normalize_layer.cpp is introduced by this paper. You can refer to it for details.

from refinedet.

I have read this paper, the paper use standard L2-norm.

The code in normalize_layer.cpp is different from the paper.

from refinedet.

@auroua
The normalize_layer.cpp is the standard L2-norm, but it does L2-norm per each pixel in a feature map instead of the whole. Just as the paper says:

from refinedet.

If the feature map is [batch_size, 5, 5, 3], and the three channels is [x, y, z].
In position [0, 0] , the normalize should be
[x[0, 0], y[0, 0], z[0, 0]]/math.sqrt(x[0, 0]^2 + y[0, 0]^2 + z[0, 0]^2).
but the code is
[x[0, 0], y[0, 0], z[0, 0]]/math.sqrt(x[0, 0] + y[0, 0] + z[0, 0]).
I don't know why the code in normalize_layer.cpp is this form.

from refinedet.

@auroua
The line 101 'caffe_sqr(dim, bottom_data, buffer_data)' is for x[0, 0]^2, y[0, 0]^2, z[0, 0]^2.

from refinedet.

@sfzhang15
😰 😰
Thanks

from refinedet.