Computer Vision¶
working pillar 9
Image fundamentals through modern detection, with an eye to scientific imagery. An active growth area.
Pages¶
| Page | Covers |
|---|---|
| Image fundamentals | Sampling and aliasing, filtering, gamma, colour, the pinhole model, interpolation |
| Classical computer vision | Features, RANSAC, epipolar geometry, stereo depth error, optical flow, calibration |
| Detection and segmentation | The four tasks and their metrics, NMS, focal loss, anchor and set-prediction families |
| Vision-language models | Contrastive alignment, generative VLM structure, what they cannot do |
| Scientific and remote-sensing imagery | Radiometry, spectral indices, georeferencing, spatial CV, fire detection |
The theme¶
Deep learning replaced the matching, not the geometry. Which pixel corresponds to which is now learned, and learned well. Triangulation, calibration, the epipolar constraint and the \(Z^2\) depth-error law are unchanged, because they are statements about projection rather than about appearance.
Most production systems are therefore hybrids: learn the correspondences, solve the geometry classically. Knowing which half of a pipeline is which is what lets you diagnose it — a depth error that scales as \(Z^2\) is geometry, and no amount of retraining will move it.
A note for the scientific case¶
Scientific imagery is the page that departs most from standard CV practice, because the pixels are measurements with units. Normalising to 8-bit, bilinear-resampling a class map, or splitting spatially-correlated pixels at random are all routine in vision work and all destroy something in a radiometric product.
Connections¶
- Deep Learning — the backbones and architectures.
- Meteorology › Fire weather — what the detections feed.
- Scientific ML — imagery as one modality among several.