Makes use of whole scene rotational normalization, similar to d6ad076f.
When building a mapping from input objects to output objects, we notice that objects in the input and output seem to stay within vertical columns. For this input, we enhanced the matching code to choose the object in that vertical column that is the most similar.