Computer vision (CV) is the engineering discipline that builds systems able to extract meaning from images and video. Where a human glances at a photo and instantly parses objects, text, faces, depth, and intent, a computer starts with nothing but a grid of numbers. The entire field is about turning that raw numeric grid into structured understanding: a class label, a set of bounding boxes, a pixel-accurate mask, a caption, a 3D pose, or a decision that drives an action. This first lesson frames the field as it stands in 2026 and hands you the map you will follow for the rest of the course.
From pixels to meaning: the core abstraction
At the lowest level, a digital image is a tensor of intensities. A grayscale image is a 2D array of shape (height, width); a standard color image is a 3D array of shape (height, width, 3) holding red, green, and blue channel values, each typically an 8-bit integer in the range 0-255. A one-minute video at 30 frames per second is simply 1,800 such tensors in sequence. Every computer vision task, no matter how sophisticated, is ultimately a function that maps this tensor to some output:
- Classification maps the tensor to one label (or a probability distribution over labels).
- Detection maps it to a list of bounding boxes, each with a label and a confidence.
- Segmentation maps it to a per-pixel label map — every pixel gets a class or an instance ID.
- Captioning / visual question answering maps it to a sequence of words.
- Pose / depth estimation maps it to keypoints or a per-pixel distance.
Keep this mental model — tensor in, structured output out — because it anchors everything that follows. When a model confuses you, ask what its input tensor is and what structure it must produce; the architecture is just the machine that connects the two.
The following table makes the mapping concrete and previews the whole course. Read it as a lookup you will return to whenever you are unsure which family of model a problem belongs to.
| Task | Input | Output structure | Typical 2026 model |
|---|---|---|---|
| Classification | One image | One label / probability vector | CNN (ResNet, ConvNeXt) or ViT |
| Object detection | One image | List of (box, class, score) | YOLO, DETR, RT-DETR |
| Semantic segmentation | One image | Per-pixel class map | DeepLab, Mask2Former, SAM |
| Instance/panoptic segmentation | One image | Per-object masks + classes | Mask2Former, YOLO-seg |
| Keypoint / pose | One image | Coordinates of joints/landmarks | YOLO-pose, ViTPose |
| Depth / 3D | One or more images | Per-pixel distance / point cloud | Depth Anything-style monocular models |
| Captioning / VQA | Image + optional text | Text answer | GPT-6.2 Sol, Gemini 4 Argon, Claude vision |
| Generation | Text (or image) prompt | New image | Diffusion models |
Notice that the input tensor is nearly always the same grid of numbers; what changes is the structure the model is asked to produce and, therefore, the head and training objective bolted onto a shared backbone. This is why a strong image backbone is reusable across so many tasks — a theme we return to with transfer learning.