Computer vision: tasks and methods

In short: Computer vision lets computers understand images. Typical tasks: classification (what is shown?), object detection (where?), segmentation (which pixels?), face recognition, text recognition (OCR) and depth estimation.

In more detail: Each task has its own outputs and loss functions but uses similar building blocks (CNN or vision transformer).

In detail

• Classification: one label per image (“cat”). Data set ImageNet, 1000 classes.
• Object detection: boxes (bounding boxes) and classes: YOLO (fast), Faster R-CNN, DETR.
• Segmentation: pixel-accurate masks. Semantic (class per pixel), instance (each object individually). Segment Anything (SAM).
• Pose estimation: joint points of people.
• OCR: text from images (Tesseract, modern models). Documents, receipts.
• Depth and 3D: estimate distance from one image, stereo, NeRF/Gaussian splatting.
• Applications: medicine (X-ray), cars, quality control, barcode scanner in phones, camera functions.
• Data augmentation: flip, rotate images, change brightness so the model becomes more robust.
• Privacy: face recognition is legally restricted, see EU AI Act and rules for AI.

See also: Convolutional networks (CNN) for images, Vision transformer and CLIP