Semantic Color Transfer -- more examples and failure cases
Jump to failure cases.
We show two examples per keyword (indicated at the top of each section).
gold
input
|
output
|
input
|
output
|
autumn
input
|
output
|
input
|
output
|
sky
input
|
output
|
input
|
output
|
strawberry
input
|
output
|
input
|
output
|
grass
input
|
output
|
input
|
output
|
banana
input
|
output
|
input
|
output
|
sunset
input
|
output
|
input
|
output
|
Failure cases
autumn
input
|
output
|
The grass in the forground is too brown as grass usually does not wilt like this. The reason for this failure case is that there are very few autumn colors in the input image, which makes it difficult to estimate the regions to process. This becomes clear when looking at the weight map:
strawberry
input
|
output
|
This image is annotated with strawberry even though there is none. It is possible that the there is a different reason why the scene is related to it, e.g. the person and the photographer just ate strawberries. In any case, the algorithm processes all image regions that are closest to strawberry and increases the red content. This affects the skin, the dress and a bit of the floor.
grass
input
|
output
|
The result can look unnatural if the processing is very strong.