Flashcard: Advanced Image Recognition and Classification — 20 carte

Interest point detection — purpose?

Identify repeatable, distinctive features in images.

Harris detector — key idea?

Detect corners via intensity autocorrelation analysis.

Scale-adapted Harris — extension?

Detects features across multiple scales.

Laplacian-based detector — used for?

Blob detection using LoG or DoG.

SIFT — developed by?

Lowe in 2004.

Matching algorithm — role?

Establish correspondences between features.

Feature descriptors — examples?

SIFT, SURF, learned CNN features.

Hand-crafted features — definition?

Manually designed to encode image properties.

Learned features — obtained how?

Automatically learned via neural networks.

Image classification — task?

Assign label to entire image.

Class scores — meaning?

Confidence levels for each class.

Datasets — examples?

MNIST, ImageNet.

Learning paradigms — types?

Supervised, unsupervised, semi-supervised.

Supervised learning — data?

Labeled input-output pairs.

Semi-supervised learning — data?

Labeled plus unlabeled data.

Linear classifier — decision boundary?

A hyperplane in feature space.

Hyperplane equation — in 2D?

w₁x₁ + w₂x₂ + b = 0.

Support vectors — what?

Closest points defining the margin.

Maximum margin — goal?

Maximize distance between hyperplane and support vectors.

Slack variables — purpose?

Handle non-separable data with soft margin.