Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials

multiclass image segmentation

• common approach: MAP inference in a CRF
  • inability to model long-range connections
• fc crf: slow
  • contribution: efficient inference algorithm: O(N) approximation
  • downsample, convolution, then upsample
• refer to paper for more details

I need to learn signal processing now…