$\color{black}\rule{365px}{3px}$

Link to Paper:

“Fully Convolutional Networks for Semantic Segmentation” - 2015

arxiv.org

Table of Contents

1. Introduction

$\color{black}\rule{365px}{3px}$

Jonathan Long | Evan Shelhamer | Trevor Darrell

Significance of the Paper

• Pioneering Fully Convolutional Approach:

  Introduced the concept of Fully Convolutional Networks (FCNs), which replace fully connected layers with convolutional layers, enabling end-to-end dense predictions for pixel-level tasks like semantic segmentation.

• The first work to train FCNs end-to-end:

  • (1) for pixelwise prediction and
  • (2) from supervised pre-training

• Skip Connections for Multi-Scale Fusion:

  Introduced skip connections to combine high-level semantic information from deeper layers with low-level spatial details from shallower layers, progressively refining predictions (e.g., FCN-32s, FCN-16s, FCN-8s).