Faster R-CNN: Training

RPN

  • Training

    • It randomly sample equal sized positive / negative boxes (totally ~ \(256\))

    • pre-trained backbone are also turned

  • normalization not required and could be simplified

  • weights initialization: from a normal distribution \(N(0, 0.01)\)

    \[\mathcal{N}(0, \, 0.01)\]
  • pre-trained backbone turning

    • ZF: all layers

    • VGG: \(conv3_1\) and up

Overall Process

4-step alternating training

  1. train RPN above

  2. train detector, initialized by

    • pre-trained backbone

    • proposals of RPN trained in step 1

    • backbone not shared between RPN and detector

  3. TBD: fine tune RPN, fix backbone

  4. fine tune detector, fix backbone

Back to Object Detection.