stem/AI/Neural Networks/CNN/FCN
andy 1c441487f9 vault backup: 2023-06-02 17:17:29
Affected files:
.obsidian/workspace-mobile.json
.obsidian/workspace.json
Lab/Scratch Domain.md
Money/Econ.md
STEM/AI/Classification/Classification.md
STEM/AI/Classification/README.md
STEM/AI/Classification/Supervised.md
STEM/AI/Neural Networks/CNN/Examples.md
STEM/AI/Neural Networks/CNN/FCN/FCN.md
STEM/AI/Neural Networks/CNN/FCN/FlowNet.md
STEM/AI/Neural Networks/CV/Filters.md
STEM/img/coordinate-change.png
STEM/img/gaussian-class.png
Tattoo/Engineering.md
Want.md
2023-06-02 17:17:29 +01:00
..
FCN.md vault backup: 2023-06-02 17:17:29 2023-06-02 17:17:29 +01:00
FlowNet.md vault backup: 2023-06-02 17:17:29 2023-06-02 17:17:29 +01:00
Highway Networks.md vault backup: 2023-05-27 23:02:51 2023-05-27 23:02:51 +01:00
README.md vault backup: 2023-05-31 21:29:04 2023-05-31 21:29:04 +01:00
ResNet.md vault backup: 2023-05-31 17:33:05 2023-05-31 17:33:05 +01:00
Skip Connections.md vault backup: 2023-05-27 23:02:51 2023-05-27 23:02:51 +01:00
Super-Resolution.md vault backup: 2023-05-26 18:52:08 2023-05-26 18:52:08 +01:00

Fully Convolutional Network

Convolutional Layer and UpConv with Activation Functions#ReLu but no others (pooling)

  • All some sort of Encoder-Decoder

Contractive → UpConv

Image Segmentation

  • For visual output
    • Previously image \rightarrow vector
  • Additional layers to up-sample representation to an image

!fcn-uses.png

!fcn-arch.png

Training

  • Rarely from scratch
  • Pre-trained weights
  • Replace final layers
    • MLP layers
    • White-noise initialised
  • Add upconv layer(s)
    • Fine-tune train
    • Freeze others
    • Annotated GT images
  • Can use summed per-pixel log Deep Learning#Loss Function

Evaluation

fcn-eval

  • SDS
    • Classical method
    • 52% mAP
  • FCN
    • 62% mAP
  • Intersection over Union
    • IOU
    • Jaccard
    • Averaged over all images
    • J(A,B)=\frac{|A\cap B|}{|A\cup B|}