Image Segmentation Evaluation

December 1, 2015 · View on GitHub

Martin Keršner, m.kersner@gmail.com

Evaluation metrics for image segmentation inspired by paper Fully Convolutional Networks for Semantic Segmentation.

Pixel accuracy

Mean accuracy

Mean IU

Frequency Weighted IU

Explanatory notes

  • n_cl : number of classes included in ground truth segmentation
  • n_ij : number of pixels of class i predicted to belong to class j
  • t_i : total number of pixels of class i in ground truth segmentation

Contents

  1. 0.1Pixel accuracy
  2. 0.2Mean accuracy
  3. 0.3Mean IU
  4. 0.4Frequency Weighted IU