List of errors found in the paper after publication
---------------------------------------------------

Figure 2: 

Labels are plotted wrong. We built a 5-by-5 matrix with the SORD label values, where rows were labels and columns were ranks. We inadvertedly swapped the axes when plotting the figure, and used rows as ranks and columns as labels. For instance, this can be seen in the label 'flawed', where values for ranks 1 and 3 should have the same value since they are 1 unit far from the true label, but instead rank 1 has a slightly higher value. This is simply a plotting script bug. All experiments used a correct formulation of the SORD labels.

References:

We forgot to cite the repositories of the off-the-shelf Keras implementations of Deeplabv3+ and FCN, and only cited the original scientific articles for these network architectures. We thank Emil Zakirov (@bonlime) and @aurora95 for these valuable contributions.

https://github.com/bonlime/keras-deeplab-v3-plus
https://github.com/aurora95/Keras-FCN