Notes - Fully Convolutional Networks for Semantic Segmentation

Fully Convolutional Networks for Semantic Segmentation

Our key insight is to build “fully convolutional” networks that take input of arbitrary size and produce correspondingly-sized output with efficient inference
and learning

Semantic segmentation takes images as inputs and ouputs lines, curves, shapes etc that sement the image.

The novelty of the paper is that it
1) does pixelwixe prediction with FCNs trained end to end
2) uses pretrained networks

Related Work
Fully Convolutional Networks 
- have height width depth
- receptive fields are where deeper feilds are impacted by specific areas in shallower layers
- translation invariance is that an image translation does not affect how its percieved
convnets apparently are translation invariant because they work on local regions and lok for relative relationships

Show fully convolutional networks (FCNs) trained pixel to pixel on semantic segmentaion are better than current methods.