multiple classes in mask

[anushree9699]

Hi, so I have trained the model to segment my custom dataset into multiple classes (different zones of the object). My masks are having 0, 1, 2 and 3 as pixel values to represent backgound and the 3 zones of the object. However, when I try to run predict.py it generates completely black masks, I tried to parse through the image to check if it contains any non-zero values and found out that all pixels are zeros. Is there anywhere in the training or prediction code I would have to change to accommodate my requirements? I gave the number of classes as 4 in training and prediction arguments, is that incorrect? Should I have given 3 as perhaps you don't count background as a class?

[ajithvcoder]

@anushree9699

1. Did u checked all u r train mask with any script does it contain only 0, 1, 2 and 3 as pixel values with opencv or pillow ?

Even though if you are keeping it 0, 1, 2 and 3 as pixel values and still if there is a problem i think while doing some image processing it would have got confused as the pixel range is very near. You can keep 0, 80, 170, 255 for the train masks of 4 different classes and train. Let me know if u r facing any issues even after this type of change

[ajithvcoder]

Then dont change the class keep it as 4

[anushree9699]

I used the pillow, this was the function i used to parse each image:

def parse_jpeg_image(image_path):
   try:
       with Image.open(image_path) as img:
           width, height = img.size
           mode = img.mode
           print(f"Image Size: {width}x{height}")
           print(f"Image Mode: {mode}")
           print(f"Number of Channels: {len(mode)}")

           zero_count = 0
           non_zero_counts = {}
           for y in range(height):
               for x in range(width):
                   pixel = img.getpixel((x, y))
                   if pixel == 0:
                       zero_count += 1
                   else:
                       if pixel in non_zero_counts:
                           non_zero_counts[pixel] += 1
                       else:
                           non_zero_counts[pixel] = 1

           print("Zero pixel count:", zero_count)
           if non_zero_counts:
               print("Non-zero pixel values and their counts:")
               for pixel, count in non_zero_counts.items():
                   print(f"Pixel value: {pixel}, Count: {count}")
           else:
               print("No non-zero pixel values found in the image.")
   except FileNotFoundError:
       print("File not found.")

And this was the output (giving for one image as example):
Image Size: 1920x1080
Image Mode: P
Number of Channels: 1
Zero pixel count: 1905463
Non-zero pixel values and their counts:
Pixel value: 2, Count: 132729
Pixel value: 3, Count: 22865
Pixel value: 1, Count: 12543

I will try with the values you suggested. Thank you.

[anushree9699]

I retrained the model as suggested by you by using the mask images with pixel values 0, 85, 170 and 255 and I am still getting the same result (completely black and empty mask) by using predict.

The training masks which I use are like:
Image Size: 1920x1080
Image Mode: L
Number of Channels: 1
Zero pixel count: 1949404
Non-zero pixel values and their counts:
Pixel value: 170, Count: 86534
Pixel value: 255, Count: 10566
Pixel value: 85, Count: 27096

And the mask the predict.py script creates are like:
Image Size: 1920x1080
Image Mode: L
Number of Channels: 1
Zero pixel count: 2073600
No non-zero pixel values found in the image.

Can there be something else going wrong?

[ajithvcoder]

sure what is the validation/testing accuracy ? what is the training accuracy.
if its possible to share the code/notebook kindly share it to inocajith21.5@gmail.com i will look into it.