🟢 1. Semantic Segmentation

• Semantic Segmentation aims to predict a category for each pixel in an image.
• Example: An image with 2 dogs, 1 cat, and a wolf → output a pixel-wise mask with class labels.
• Common in autonomous driving, medical imaging, and scene understanding.

🪟 2. Sliding Window in Segmentation

• Idea: Slide a window across the image → classify the central pixel.
• ✅ Works in theory, but:
  • ❌ Super computationally expensive.
  • Many redundant computations.
  • Doesn’t scale well to high-res images.

🔄 3. Fully Convolutional Networks (FCN)

• FCNs solve this by removing fully connected layers.
• They apply convolution and upsampling across the entire image.
• ✅ Output: Dense pixel-wise predictions.
• Faster and more efficient than sliding windows.

⬆️ 4. Upsampling & Unpooling

✅ Upsampling

• Increases the spatial resolution of feature maps.
• Examples: Displaying low-res images on high-res screens.