Flashcards
ADL2021
Flashcards do conjunto (10)
- 1What improvements are introduced by the Style GAN ?1. Mapping the latent vector z on a vector w using a pipeline of fully connected layers.
2. Information from w is fed in on multiple levels instead of being the "source" of the upconvolution
3. Additionly to style information noise is added - 2What is the reasoning behind the computation of the vector w?The goal is an disentangled latent vector. That is a vector that encodes every variation in the resulting image in only one of its dimensions. Since some combinations of attributes are missing the resulting z vector is distorted/warped; the intention of the Fully connected network f(z) -> w is to undo this warping
- 3Why is noise added to the image?To account for stochastic features; such as the texture of skin, hairs, freckles, etc...
- 4How is the style GAN related to Style Transfer?Using AdaIN to mix style information on multiple levels; Instead of extracting the style from another image, the style is extracted from the latent vector w using learned Affine transforms (i.e. Dense Layer)
- 5How is the Style GAN trained?Progressive Growing: The full Model is not trained from scratch; instead more layers are added (to the generator & Discriminator) to increase the resolution of resulting images progressivley;
To avoid shocks to well trained low res layers when increasing the resolution, new layers are added smoothly by interpolating between them and a simple upsampling function - 6What is AdaIN?Adaptive Instance Normalisation: Used to mix the style information into an image.
General: Add statistical information from one distribution to another by first removing the "old" ones and then adding the new.
Remove Mean and Std from image, then add Style infromation - 7How can a (Style) GAN be used to edit images?1. Find the latent code of the image that should be edited
2. Change the latent code (systematically) to change certain features of the image - 8How can the latent vector for a given image be computed?1. Naive: compute L2 loss of randomly generated image and target image, and minimize using gradient descent on the latent code; gets stuck in local minima
2. improved: Use e.g. VGG to extract another latent code and minimize L2 loss of the latent VGG codes instead (very slow)
3. Start from a better initial guess by training a second encoder - 9What is the difference between StyleGAN and StyleGAN3?StyleGan has the problem that features are glued to certain pixel coordinates instead of on surfaces of the depicted objects.
StyleGAN3 tackles this problem by representign the image as a 2dimensional wavefunction instead of grid of discrete values. The Wavefunction makes interpolation possible - 10What is YOLO?Stands for You only look once. Model for Object detection.
detects globally instead of predicting probabilities for parts of the image.
The prediction is split into a grid; for every grid cell multiple bounding boxes are predicted as well as one "set" of class probabilities;
Informações
- Descrição
- FlashCards sobre ADL2021, criado por Laurenz Nippelbeißer em 04-04-2022.
- Criado por
- Laurenz Nippelbeißer
- Visualizações
- 1
- Avaliação