Semantic Segmentation with Fully Convolutional Networks (FCN-8s)

This repository contains an implementation and analysis of semantic segmentation using Fully Convolutional Networks (FCNs), following the seminal paper by Long et al. (2015).

The project focuses on understanding how classification CNN architectures can be transformed into dense, pixel-wise prediction models through fully convolutional design and skip connections.

---

Reference Paper

• Jonathan Long, Evan Shelhamer, Trevor Darrell
  Fully Convolutional Networks for Semantic Segmentation
  https://arxiv.org/abs/1411.4038

---

Problem Setting

• Task: Semantic Segmentation
• Goal: Assign a semantic label to every pixel in an image
• Challenge:
  • Preserve fine spatial details
  • Combine high-level semantic information with low-level appearance cues

---

Model Architecture

• Fully Convolutional Network (FCN)
  • Converted a classification CNN into a fully convolutional model
  • Replaced fully connected layers with convolution layers
• FCN-8s architecture
  • Combines predictions from:
    • Deep, coarse layers (semantic information)
    • Shallow, fine layers (spatial detail)
  • Uses skip connections and in-network upsampling
• Upsampling:
  • Learnable deconvolution layers for pixel-level prediction

---

Key Takeaways

• Fully convolutional networks enable efficient, end-to-end semantic segmentation
• Skip connections are critical for recovering fine-grained spatial details
• Semantic segmentation requires balancing global context and local precision