Learning to see

Classifying images with features and nearest neighbours, convolutional neural networks, and detection and segmentation with IoU, NMS and average precision.

About 5 hours of study

Learning objectives

Lessons

  1. 1

    Classifying images with features and nearest neighbours Intermediate

    Teach a program to name what is in a picture. Treat images as vectors, classify them with k-nearest neighbours, see why raw pixels fail when objects move, and build HOG-style features from gradient orientations that make classification robust.

  2. 2

    Convolutional neural networks for images Advanced

    Let the network learn its own kernels. Follow an image through convolution layers with many channels, ReLU and max pooling to a classifier, compute output shapes and parameter counts, and see why weight sharing and pooling suit images.

  3. 3

    Object detection metrics and segmentation Advanced

    Go from "what is in the image" to "where is it". Measure box overlap with IoU, remove duplicate detections with non-maximum suppression, score a detector with precision, recall and average precision, and evaluate segmentation masks with IoU and Dice.

Module practice

Module quiz

1. What does a k-NN classifier do with a new image?
2. What is the IoU of two identical boxes?
3. Why do HOG features tolerate small shifts of an object?
4. A conv layer has 32 filters of 3 x 3 on 16 input channels. How many parameters does it have?
5. Why can a duplicate detection of an already-found object not count as a true positive?
6. A mask pair has IoU 0.5. What is its Dice coefficient?

‹ Geometry and matchingFinal assessment ›