Short & Breezy Flashcards
Covering only areas that I couldn't achieve through the use of GPT flashcards (17 cards)
What does Chan-vese have as an advantage over the Snake model?
- Requires no parameterisation
- Computationally efficient
What does the data term refer to in the Snake model?
The data term refers to the cost of assigning a specific label to a pixel based on observed features, like intensity.
What does the smoothness term refer to in the Snake model?
Smoothness term refers to the penalties applied to adjacent pixels when their labels don’t align with each other. There’s less penalty applied if the pixels adjacent contain the same labels/similar labels.
What are 1x1 convolution layers used for?
They are primarily used for dimensionality reduction, enabling more efficient computation in the long run
What is the role of homogenous coordinates in projecting 3D coordinates onto a 2D plane?
They allow projection to be expressed as a linear transformation using matrix multiplication, essential for 3D to 2D mapping
What are the three separate terms for the Snake Model?
- EInternal
- EImage
- EConstraint
What does the EInternal state govern in the Snake Model?
- Continuity
- Curvature
What does the EImage state govern in the Snake model?
- Pulls contour toward features like edges
What does the EConstraint state govern in the Snake Model?
- External forces like user keypoints
What is the mathematical method for Otsu’s Thresholding?
Minimises within-class variance (weighted sum of variances of foreground/background), using variables like class probabilities and means
How does a U-net use skip connections?
Comines low-level spatial information with high-level features.
How does a Monte Carlo Markov Chain work?
- Each particle is a hypothesis surrounding the object’s state
- Each has a weight, indicating how confident that guess is
- At each frame, predict where the object might move next, update the weights based on how well each guess matches the image, and keep good guesses, discarding bad ones, and generating new ones near the good ones.
Why is Monte Carlo Markov Chain good?
- More efficient sampling
- Generates better samples the longer it’s left to iterate for
- Can handle multiple object positions at once
- Can track object/s through clutter, noise and occlusion
How does Snake work?
Aims to fit points to a shape by minimising an energy function. The energy function is defined as:
- ESnake = (alpha * EInternal) + (beta * EImage) + (gamma * EConstraint)
How is the EInternal state defined?
EInternal = EContinuity + ECurve
Where:
EContinuity = Makes sure points aren’t space unevenly
ECurve = Prevents sharp corners, encourages smoother curves
What does the EImage aspect of the energy function do?
It pulls the snake toward edges in the image
What does the EConstraint aspect of the energy function do?
It’s an optional user input or constraints, such as points the snake must pass through
