Learning Deconv_2015

19 Nov 2018

Summary

• VGG16 + Deconvolutional Network (The mirrored VGG16)
• Dense pixel-wise class probability map is obtained.
• unpooling capture example-specific structures
  • reconstruct the detailed structure of an object in finer resolution
• Learned filters capture class-specific shapes
  • deconvolutions, the activations closely related to the target classes are amplified while noisy activations from other regions are suppressed.

Training

• Batch normalization
• 2-stage traning :
  • easy examples (crop object instances using ground-truth annotations, so that an object is centered at the cropped bounding box)
  • fine-tune with more challenging examples (object proposals: Candidate proposals sufficiently overlapped with ground-truth segmentations are selected for training)
    • Generate sufficient number of candidate proposals
    • Apply trained network to obtain semantic segmentation maps of individual proposals
    • Aggregate the output of all proposals

Inference

• Employ edge-box to generate object proposals.