VL IX - CNN Architexture

LeNet Activation function

Tanh/sigmoid

Common performance metrics

• Top 1 Score

• Top 5 score

• Top 5 error (reverse of top 5 score)

Where are most parameters in a CNN

• in the FC layer

Activation Function AlexNet

ReLU

What does VGGNet change

• Striving for simplicity

• 3×3 filters with stride 1

• Pooling: 2×2 with stride 2

Structure VGGNet

Conv → Pool → Conv → Pool → … → FC

Which VGGNet is best and why

VGG-16, due to with increasing layers we get vanishing gradients due to flooting point inaccuracy

Why use Skip Connections

• Improving gradient flow

• Preserving information

• Making optimization easier

Which convolution is used in skip connection

Same convolution, otherwise conversion with zero padding or matrix of learned weights

Why do ResNets Work

• Identity is easy for the residual block to learn

• Guaranteed it will not hurt performance

Why use 1×1 Convolutions

• Use it to shrink the number of channels

• Further adds a non-linearity → learn more complex functions

Whats an Inception layer

• Layer with multiple different convolutions going to 1 output

Why use inception layers

• let the network decide which convolution works best

Why use 1×1 convolution in Inception layers

• Reduction of multiplications

Whats Xception Net

• Applying depthwise seperable convolutions

• conv layers structured into several modules with skip connections

• Filters are only applied at a certain depth

Why use xception nets

reduction of multiplications → less computations

Types of Upsampling

• Interpolation

• Transposed Conv

Left side of U-Net

Contraction Path

What does the contraciton Path do

• Captures context of the image

• Follows architecture of CNN

Right side of U-net

Expansion path (decoder)

What does the expansion path do

• Upsampling to recover spatial locations for assigning class labels to each pixel

• width up, depth down

How does the feature map change the deeper we go in a CNN

The Width goes down, but the depth goes up

Challenges detection

• convolution are translationally equivariant

  • same filter slides across all positions

  • shifting input image → output shifts by same amount

• Need for detecting multiple objects

• Identifying what makes where difficult

What does R-CNN stand for

Region-Based CNNs

What is R-CNN

• Fixed number of boxes are run over an image

• Boxes hopefully extract promising feature

• Each box has to run its features through a CNN

Difference Fast R-CNN to R-CNN

• Run CNN for entire image once

• Use region of interest pooling to extract fixed size feature maps for each region

Whats Region of interest Pooling

• Rescale variable-sized (dimension) feature maps to a fixed sized output using max pooling

How does Faster R-CNN work

Region proposal network RPN slides over feature map and dedicates if interesting feature is present

Whats an Anchor (Detection)

Potential bounding box where an object can be detected

difference Faster R-CNN and Mask R-CNN

After bounding boxes are found additionally it is determined if the pixel in the bounding box belongs to the object

Whats instance segmentation

Pixel wise mask inside of a bounding box