COMP3204: Computer Vision

What the module covered

This module covered image formation and sampling, spatial and frequency-domain filtering, edge and shape detection, segmentation, and feature-based recognition.

For the first coursework, I implemented convolution and Gaussian filtering in NumPy to create hybrid images, combining the low-frequency content of one image with the high-frequency detail of another. The implementation used vectorised array operations and separable Gaussian filters.

The group scene-recognition coursework compared classifiers for 15 categories. The pipelines included tiny images with nearest neighbours, a bag-of-visual-words representation with linear classifiers, and dense SIFT features with spatial pooling and a kernel approximation.

What I took from it

Implementing the filters made their behaviour more tangible, including how padding affects image boundaries and how a separable kernel changes the computation. The recognition work connected image representations to classification: changing the features and their normalisation was an important part of the pipeline, alongside choosing a classifier.