Computer Vision

Computer Vision

terms

• Sampling: extra sampling
• Activition function
• Loss function

• Transfer learning (e.g. already use parameters trained on Resnet)

  • Sometimes early layers are freezed
  • fastai also has a gradually increasing learning rate (as layers are closer to features)
• Global avarage pooling.

• Reduce activiation map size by

  • a) Max Polling (formerly used)
  • b) Stride (schrittweite der convolution), e.g. stride size of 2 -> Don’t do convolution on every pixel, but jump and delete jumped over pixels.
• Bottleneck feature -> Last step of feature extraction before classifier (nach global average pooling)
• Random forest (forest of randomized decision trees) nach CNN. —> Other open cv data.

• Precision and Recall

  • False Positive, False Negative

• Confused (deutsch “Verwechslung”)

  • Confusion matrix lists pairs of most confusion aka wrong classifications.
• the easy split —> what is easy to separate. -> can lead to overfitting.
• matplotlib vs. plotly

Data Preprocessing Architecture

• dataset (incl. data augmentation) + sampler —> data loader

  • image augmentation: Add different variants of the image (e.g. different crops and zooms) or different colors.
• Data Augementation: Filters, like grayscale, RGB augementation, corner augmentation etc.

Architecture of a deep learning network

First layer:

How to create explanatory images to explain what a layer is responsible for

• Images taken from the fast.ai book chapter 2 (section “What Our Image Recognizer Learned”)

First layer

• Look at the neurons with highest activations.
• The neurons are directly connected to the input image.
• Look through all the images which are highly activated for a given neuron and search for similarities.

Deeper layers

1. Check for neurons with highest activations in a given deep layer.

2. Trace back these activations to the input layer (i.e. images) they’re coming from.

   You’re asking the question: Which neurons in previous layers had the biggest impact on neurons with highest activity in this layer?

3. For re-occuring activation patterns in a given layer (a square block of 9 activations in images on the left), try to make out semantic differences in the input image (corresponding block of 9 images on the right).

Layer 2

Layer 3

Layers 4 and 5

Annotations

• Bounding box
• Instance segementation

Error in classification of arial data

1. Streets vs. Parking lots

2. Get more context!

   • Field of vision (aka. influence of pixels) Sichtweite anpassen.

   • How many pixels in previous layers influenced the specific activation map (feature map)?

     • This is usally a guassian curve -> pixels in the perifery influence the pixel less than pixels close to it.
3. Pools vs. trampolines

4. hard-negative mining

   • sample cases in which neural network has a hard job and let them be labeled.
   • Get new data which is useful for you.
5. relabeling —> new pseudo-classes

6. over-sampling: Create new image data by randomly copying different parts of the image to create new ones.

   e.g. Filter for zebra stripes.

   [
       -1, -1, 1
       -1, -1, 1
       -1, -1, 1
   ]

7. ReLu activiation function is non-linear

   • if linear, the whole Neural network would not be helpful, because I could just create another linear function f3 where f3 = f2(f1(x)) as a linear combination.

Images used for non-image tasks

e.g. audio processing

See fast.ai book section Image Recognizers Can Tackle Non-Image Tasks