A segmentation mask turns an image-level recognition problem into a pixel-level target: each input pixel receives a class or region value. The mask keeps the input image’s spatial dimensions, uses one channel, and stores integer labels rather than display colors. section 11.2.1 section 11.4
A three-class pet mask
In the Oxford-IIIT Pets data, a photograph has a matching PNG mask with the same filename. Each mask pixel stores 1 for foreground, 2 for background, or 3 for contour. Loading it in grayscale preserves these labels; preprocessing can subtract 1 to create zero-based targets. section 11.2.1
When preparing masks for a model, keep these practical rules in view:
- Pair each input image with its corresponding mask path, commonly using matching filenames in separate image and annotation directories. section 11.2.1
- Use a single-channel integer array for the target, such as a batch shaped for height, width, and one channel. section 11.2.1
- Treat mask values as labels, not ordinary image intensities; visualization may remap them to contrasting colors for inspection. section 11.2.1
Masks can also describe individual objects rather than fixed semantic classes. Segment Anything accepts an image with a point or box prompt and predicts candidate masks for the indicated object, which can then be compared with a ground-truth mask. section 11.3.2 section 11.3.4
def display_target(target_array): normalized_array = (target_array.astype("uint8") - 1) * 127 #1 plt.axis("off") plt.imshow(normalized_array[:, :, 0]) img = img_to_array(load_img(target_paths[9], color_mode="grayscale")) #2 display_target(img)
Inspect a target mask and its label encoding with a small grayscale-display utility. section 11.2.1
import numpy as np import random img_size = (200, 200) #1 num_imgs = len(input_img_paths) #2 random.Random(1337).shuffle(input_img_paths) #3 random.Random(1337).shuffle(target_paths) #3 def path_to_input_image(path): return img_to_array(load_img(path, target_size=img_size)) def path_to_target(path): img = img_to_array( load_img(path, target_size=img_size, color_mode="grayscale") ) img = img.astype("uint8") - 1 #4 return img input_imgs = np.zeros((num_imgs,) + img_size + (3,), dtype="float32") #5 targets = np.zeros((num_imgs,) + img_size + (1,), dtype="uint8") #5 for i in range(num_imgs): #5 input_imgs[i] = path_to_input_image(input_img_paths[i]) #5 targets[i] = path_to_target(target_paths[i]) #5
Follow the NumPy preparation pipeline that resizes inputs and stores masks as single-channel integer targets. section 11.2.1
