Computer vision is getting useful information out of images and video: what is in a picture, where it is, how it moves, how far away it is. It runs in phone cameras, factories, medical imaging and robots. An image is just a grid of numbers, and the whole field is about turning that grid into meaning despite light, angle and clutter.
Key concepts
- Image formation: pixels, colour spaces, resolution, noise and the camera model that links 3D points to 2D coordinates.
- Classic processing: filtering, edges, corners, feature descriptors and geometric matching, still useful and fast.
- Core tasks: classification, detection, segmentation, tracking, pose estimation and depth.
- Learned models: convolutional and attention-based networks trained on labelled images, usually from a pretrained start.
- Geometry: calibration, stereo and structure from motion recover 3D from several views.
Vocabulary
- Convolution: sliding a small filter over an image to detect local patterns.
- Bounding box: a rectangle around a detected object.
- Intersection over union: overlap of a predicted and a true region divided by their combined area.
- Semantic and instance segmentation: labelling each pixel by class, or by each separate object.
- Optical flow: the apparent motion of pixels between two frames.
What to study, in order
- Load an image as an array, and implement blur, edge detection and thresholding by hand.
- Learn the pinhole camera model and calibrate a camera with a printed chessboard.
- Match features between two photos and estimate the transform that aligns them.
- Train a small classifier, then fine-tune a pretrained network on your own images.
- Measure detection with intersection over union, precision and recall, and inspect the failures by eye.
Related tags
#deep-learning #computer-graphics #slam #signal-processing #robotics