Computer Vision & ML Exam Essentials

Mastering computer vision and machine learning requires understanding foundational image processing, core ML algorithms, advanced deep learning architectures for visual tasks, generative modeling for synthesis, and 3D vision principles for spatial and temporal data.

Core Principles

  • Image Filtering: Kernels like 'impulse' (no change), 'translated impulse' (shift), and 'box filter' (blur) perform neighborhood operations via cross-correlation, while sharpening accentuates local differences.
  • Fourier Transforms: The Discrete Fourier Transform (DFT) decomposes signals into frequency components, revealing properties like translation affecting phase, scaling inversely affecting frequency content, and rotation preserving orientation in the frequency domain.
  • Machine Learning Generalization: Models must balance fitting training data and generalizing to unseen data; underfitting implies a too-simple model, while overfitting indicates excessive complexity, often mitigated by regularization (e.g., L2 norm with parameter lambda).
  • Gradient Descent Optimization: Iteratively adjust model parameters (theta) in the direction of steepest descent of the loss function J(theta), with the learning rate (eta_t) controlling step size, often decreasing over time for better convergence.
  • Deep Learning Architectures: Convolutional Neural Networks (CNNs) use hierarchical feature extraction with layers like convolution, ReLU, and pooling; advanced models like AlexNet (2012), VGG (2014), and ResNet (2016) progressively reduced ImageNet classification error.
  • Generative Adversarial Networks (GANs): A Generator (G) synthesizes fake data to fool a Discriminator (D), which learns to distinguish real from fake, creating a minimax game that ideally converges when G reproduces the data distribution.
  • Vision Transformers (ViT): Split images into patches, linearly project them into tokens, and feed them into a standard transformer encoder, leveraging attention mechanisms and positional encodings to capture global relationships.
  • 3D Vision & Motion: Processing video involves extending 2D vision with temporal components (e.g., 3D CNNs, 3D Vision Transformers, RNNs), while geometric modeling uses concepts like homogeneous coordinates and projection matrices for 3D-to-2D mapping.

Formulas

  • Discrete Fourier Transform (DFT): $$F[u] = \sum_{n=0}^{N-1} f[n] \exp \left(-2\pi i \frac{un}{N}\right)$$
  • Least Squares Objective (Loss): $$\mathcal{L}(\hat{y}, y) = (\hat{y} - y)^2 \quad \text{where } \hat{y} = f_\theta(x)$$
  • Softmax Regression (Probability Density Function): $$\hat{y}_j = \frac{e^{z_j}}{\sum_{k=1}^K e^{z_k}}$$
  • Gradient Descent Update: $$\theta^{t+1} = \theta^t - \eta_t \nabla_\theta J(\theta) \Big|_{\theta=\theta^t}$$
  • Instance Normalization: $$\hat{h}_k = \frac{h_k - E[h_k]}{\sqrt{Var[h_k]}}$$
  • Generative Adversarial Network (GAN) Objective: $$\min_G \max_D E_{x,y}[\log D(x, G(x))] + \log(1 - D(x, y))$$

Key Terms

  • Cross-correlation: An image filtering operation where a kernel slides over an image, computing a weighted sum of neighborhood pixels to produce an output pixel, used for effects like blurring or sharpening.
  • K-Nearest Neighbor (K-NN): A non-parametric classification algorithm that assigns a label to an input based on the most common label among its 'k' closest examples in the training data.
  • Least Squares Objective: A loss function that minimizes the sum of squared differences between predicted values (model output) and actual target values, commonly used in linear regression.
  • Softmax Regression: A generalization of logistic regression to multiple classes, producing a vector of probabilities for each class that sum to one, effectively a probability density function.
  • Backpropagation: An algorithm for efficiently computing the gradients of the loss function with respect to all network parameters by applying the chain rule in reverse through the computation graph.
  • Stochastic Gradient Descent (SGD): An optimization algorithm that estimates the gradient of the loss function using a small random subset of training examples (minibatch) rather than the entire dataset, speeding up training.
  • Instance Normalization: Normalizes the activations of a single hidden unit to have mean 0 and standard deviation 1, estimated over each individual input instance, improving training stability.
  • Generative Adversarial Network (GAN): A framework where two neural networks, a Generator and a Discriminator, compete in a zero-sum game to learn to generate realistic data.
  • Positional Encoding: Location information added to token code vectors in transformer models to account for the order or spatial arrangement of input elements, as transformers are permutation-equivariant by default.
  • Homogeneous Coordinates: An extended coordinate system (e.g., adding a 'w' coordinate to 2D/3D points) that simplifies geometric transformations like translation, scaling, and rotation into matrix multiplications.

Pitfalls to Avoid

  • Underfitting: Using a model too simple (e.g., K=1 for k-NN) fails to capture underlying data patterns, resulting in poor performance on both training and test data.
  • Overfitting: Employing an overly complex model (e.g., K=10 for k-NN) fits noise in the training data, leading to excellent training performance but poor generalization to new, unseen data.
  • Ignoring Learning Rate Schedules: Maintaining a constant high learning rate (eta_t) can cause oscillations around the optimum, preventing convergence, while a too-low rate leads to extremely slow training.
  • Skipping Vectorization: Processing data examples one-by-one in deep learning (e.g., batch processing without vectorized code) is computationally inefficient, significantly slowing down training.
  • Using Averaging for Motion Analysis: Simple averaging models for video tasks (e.g., action recognition) cannot capture or analyze motion, limiting their effectiveness for dynamic events.

Related cheat sheets

  • EECS 442: Computer Vision & ML Fundamentals
  • Heat Transfer Solutions: Core Principles & Pitfalls
  • Partial Differential Equations: A Cheat Sheet
  • Experimental Probability Cheat Sheet
  • Mastering Linear Graphs and Rules
  • Probability Basics Cheat Sheet

More like this

  • Science & Academic cheat sheets
  • Explore all cheat sheets

ClipSheet — AI Cheat Sheet Generator

ClipSheet transforms YouTube videos, PDFs, and text into structured cheat sheets and study notes using AI. Built for students, professionals, and content creators who need to learn faster.

Features

  • AI-powered extraction of key concepts, formulas, and action steps
  • Automatic quiz and flashcard generation for active recall
  • PDF export and public sharing via unique URLs
  • Support for YouTube videos, PDFs, and raw text input

Browse by Category

  • All Cheat Sheets
  • Science & Academic
  • Technology
  • Health & Fitness
  • Coding
  • Business
  • Education
  • Productivity
  • Finance
  • Lifestyle

Legal

  • Privacy Policy
  • Terms of Service
  • Imprint
  • Verträge hier kündigen