Setting particular pixel to specified color in the output of segmentation map

Python
1

I was able to run semantic segmentation on the below image.The problem is given all pixels belonging to the sky category I need to set them to white
.
For this my approach is as follows

  1. obtain semantic map output
  2. then find all pixels belonging to the sky class[which in this case have their color value set to 6,230,230]
  3. And then set them all to 255.

The original image

ADE_val_00000004
ADE_val_00000004683×512 45.2 KB

And its semantic map output is as follows
res_img

This is what i have tried

indice_r=np.where(image[:,:,0]==6)
indice_g=np.where(image[:,:,1]==230)
indice_b=np.where(image[:,:,2]==230)
#Set the pixel value for those found indices(here we are performing this across B ) to 0
#The same step is done twice for R and G respectively
for i in range(len(indice_b[0])):
  for j in range(len(indice_b[1])):
    new_img[i,j,2]=255

Unfortunately this approach didn’t work as it only made a small section on the top left white
So i was wondering is this really the right approach and how can I generalize this to handle any class rather than sky.For example what if we want to set the pixels belonging to the person to white.

Note in all images the sky class has been color coded as [6,230,230 This is the standard choice of the semantic pallete chosen in the ADE dataset.

Another approach is to use bitwise_and to select the blue regions in the semantic map and then subtract that from the original image.However that requires a lower and upper range which may differ from image to image.

do you think the bitwise_and would help?

2

In case it’s not clear why my first approach to setting pixels to white isn’t working.
this is the result for the first approach
wrong_result

So i tried using the bitwise idea.
Here i converted the image to hsv to get a better understanding of which pixels are set to blue or light blue
The HSV plot is shown below
hsv_ade
from this i think the light blue lower and upper ranges are
(77,90,0) and (125,250,225)
Using this i created the bitwise mask and obtained this beauty
code for masking
bitwise_mask
Code for masking

mask = cv2.inRange(hsv_img, light_b, high_b)
result = cv2.bitwise_and(img, img, mask=mask)
plt.subplot(1, 2, 1)
plt.imshow(mask, cmap="gray")
plt.subplot(1, 2, 2)
plt.imshow(result)
plt.show()

Not sure why the sky is jet black here.

Would appreciate it if some kind soul could help me.Chris i am looking at you as you have always helped me out.

3

unclear, how you obtain this, and what you get there, exactly, please explain !

shouldn’t there be binary, per-class maps ?
imo, trying to parse an already ‘rendered’ result mapping (like your 2nd image) is a bad idea.

4

Ok so the second image was obtained by running a semantic segmentation algorithm on the first image(ADE_val0004) .this was done using the below link

As for the second part of your answer,if I wanted to get the binary mask of only the sky and then set all pixels corresponding to the sky ,how would I go about that.
thanks

5

you’ll have to dive inside the demo code there, and access the data, that generated your segmentation image

as it seems, you can get both dense pixel masks as well as outline polygons from the network prediction output: