Segmentation image is always gray? about spade HOT 3 CLOSED

If the semantic map is in color, you would first have to map the pixels values into class label numbers.

For example, if red is person, you will have to first change the pixel value of [255, 0, 0] to the class label of person.

As an alternative, if you don't have the class numbers that correspond to the RGB colors of the semantic layout, you can try modifying the code to directly learn meaningful feature embedding from the RGB values. In detail, you will comment out the code to create the one-hot vector and treat the 3-channel RGB values as input to SPADE. Our code currently does not support that.

from spade.

Did you try to insert the segmentation image instead of using one-hot vector?
just do a simple normalization.

For example, if you have a segmentation image in the shape of [256, 256, 3],
Simply an attempt to a network with a value between -1 and 1.

from spade.

Knowing that the labels are discrete value, fitting them in a continuous space of color values is likely sub-optimal.

from spade.