(Unable to) reproduce figure 1

[prhbrt]

Could you provide the input image, code, prototxt, and caffemodel to reproduce figure 1 of the paper?

I tried to reproduce figure 1 of the paper based on the pretrained model found in the readme, that is the edge maps of this figure.

However, when I do so, the result is significantly worse as the ones shown in the paper. In fact, they do not seem anything better than some Sobel-like filter. Morover, there is a 35-pixel top-left padding in the output that is awkward and mentioned in several other issues as well.

My code is based on the example notebook . My caffe version is 1.0.0.

from matplotlib import pyplot
from skimage.io import imread
import caffe
import numpy

image = imread('paper-example.jpg')
def preprocess_input(X):
    X = numpy.array(X, dtype=numpy.float32)
    X = X[...,::-1]
    X -= numpy.array([[[[104.00698793,116.66876762,122.67891434]]]])
    return X.transpose(0, 3, 1, 2) / 128

images = image[None] # images is a 1 x height x width x 3 ndarray
preprocessed = preprocess_input(images) # preprocessed is a 1 x 3 x height x width ndarray

net = caffe.Net(
    'deploy.prototxt',
    'hed_pretrained_bsds.caffemodel',
    caffe.TEST
)

net.blobs['data'].reshape(*preprocessed.shape)
net.blobs['data'].data[...] = preprocessed

net.forward()
out1 = net.blobs['sigmoid-dsn1'].data[0][0]
out2 = net.blobs['sigmoid-dsn2'].data[0][0]
out3 = net.blobs['sigmoid-dsn3'].data[0][0]
out4 = net.blobs['sigmoid-dsn4'].data[0][0]
out5 = net.blobs['sigmoid-dsn5'].data[0][0]
fuse = net.blobs['sigmoid-fuse'].data[0][0]

pyplot.rcParams['figure.figsize'] = (9, 10)
for axis, out, title in zip(
    pyplot.subplots(3,2)[1].ravel(),
    [out1, out2, out3, out4, out5, fuse],
    ['out1', 'out2', 'out3', 'out4', 'out5', 'fuse'],
):
    if out is not None:
        axis.imshow(out)
    axis.set_xticks([])
    axis.set_yticks([])
    axis.set_xlabel(title)
pyplot.tight_layout()

[prhbrt]

For easy reference, figure 1 as found in the paper.

[axhagemann]

Regarding the border check out this issue for possible fix