Transposed convolution & Fully Convolutional Neural Network

Given a kernel (e.g. 3×3 3 × 3 filter), we can get the sparse Toeplitz matrix C C whose elements are are weights in kernel.

• We can either say this kernel defines a direct convolution whose forward and backward pass are computed by C C and C⊤ C ⊤ respectively.
• We can also say this kernel defines a transposed convolution whose forward and backward pass are computed by C⊤ C ⊤ and C C respectively.

Direct Conv & Tranposed Conv

• It’s always possible to emulate (模仿) a transposed convolution with a direct conv. The disadvantage is that it usually involves adding many columns and rows of zeros to the input, resulting in a much less efficient implementation.
• Interpretation:
  
  • The simplest way to think about a transposed convolution on a given input is to imagine such an input as being the result of a direct convolution applied on some initial feature map.
  • The trasposed convolution can then be considered as the operation that allows to recover the shape of this initial feature map.
  • Note here we only recover the shape, not the exact value of the input. Anyway transposed convolution is not the inverse of convolution!!!
• To maintain the connectivity between direct conv and transposed conv, the direct conv which is used to emulate the transposed conv, may experience a specific zero padding.
• The connectivity consistency matters!!!

Fullly Convolutional Net

• only contains locally connected layers (like conv, pooling, upsampling), no dense layer used in FCC.
  
  • reduce number of paras and computation time
  • the network can work regardless of the original image size, given all connections are local.
• Segmentation net contains two path:
  
  • downsampling path: capture semantic/contextual information
  • upsampling path: recover spatial information (precise localization)
  • to further recover the spatial information we use skip connection

Skip connection:

• concatenating or summing feature maps from the downsampling path with feature maps from the upsampling path
• Merging features from various resolution levels helps combining context information with spatial information.

diff between FCN-32s, 16s, 8s

1. FCN-32 : Directly produces the segmentation map from conv7, by using a transposed convolution layer with stride 32.
2. FCN-16 : Sums the 2x upsampled prediction from_conv7_ (using a transposed convolution with stride 2) with pool4 and then produces the segmentation map, by using a transposed convolution layer with stride 16 on top of that.
3. FCN-8 : Sums the 2x upsampled conv7 (with a stride 2 transposed convolution) with_pool4_, upsamples them with a stride 2 transposed convolution and sums them with_pool3_, and applies a transposed convolution layer with stride 8 on the resulting feature maps to obtain the segmentation map.

image source

Unpooling

• Introducing a switched variable which records the location of maximum element (when using max pooling), and then unpooling the feature map

Tutorial