Computer Vision — Pipeline, Tasks, and the Terms Exams Confuse On Purpose
Computer Vision (CV) questions on CCAT often hinge on one thing: telling apart tasks that sound similar but are technically different (classification vs. detection vs. segmentation). Here's the breakdown that keeps them straight.
What Is Computer Vision?
CV enables machines to interpret visual data — images and video — and take action or make decisions based on that understanding. It relies heavily on deep learning, particularly Convolutional Neural Networks (CNNs).
The CV Pipeline
Image Capture → Preprocessing → Feature Extraction (CNN) →
Classification/Detection → Output/Action
- Preprocessing: Resizing, normalizing pixel values
- Feature Extraction: CNN layers detect edges → shapes → objects
- Classification/Detection: Final layer (e.g., Softmax) assigns labels
The Task Types That Get Confused — Master This Table
| Task | Description | Output |
|---|---|---|
| Image Classification | Assigns one label to the whole image | Class label (e.g., "cat") |
| Object Detection | Locates AND classifies multiple objects | Bounding boxes + labels |
| Semantic Segmentation | Classifies every pixel | Pixel-level map (no distinction between object instances) |
| Instance Segmentation | Segments each object instance separately | Individual object masks |
| Pose Estimation | Detects body keypoints | Skeleton coordinates |
The trap explained simply:
- Classification says "there's a cat in this image."
- Detection says "there's a cat here" (with a bounding box), and possibly other objects too.
- Semantic segmentation says "these exact pixels are 'cat' pixels" — but if there are two cats, it won't tell them apart.
- Instance segmentation does tell the two cats apart as separate objects.
This exact distinction (semantic vs. instance segmentation) is one of the most commonly tested CV nuances.
Real-World Applications — Match Them Correctly
| Domain | Application | Example |
|---|---|---|
| Content Moderation | Filtering unsafe images | Facebook, Instagram |
| Facial Recognition | Identity verification | Apple Face ID |
| Image Classification | Object identification | Google Lens |
| Object Detection | Multi-object localization | Tesla/Waymo detecting pedestrians |
| Medical Imaging | Tumor/fracture detection | Radiology AI |
| OCR | Image-to-text conversion | Google Translate camera mode |
| Augmented Reality | Digital overlay on real world | Pokémon Go, Snapchat filters |
Why CNNs Specifically?
Recall from the neural networks post: CNNs use local connections and parameter sharing, making them efficient at scanning images for local patterns (edges, textures) without needing a separate parameter for every single pixel connection — something a plain fully-connected network can't do efficiently.
Quick Recap
- CV pipeline: Capture → Preprocess → Extract Features (CNN) → Classify/Detect → Output
- Classification = one label per image
- Detection = multiple objects + bounding boxes
- Semantic segmentation = pixel-level, no instance separation
- Instance segmentation = pixel-level, WITH instance separation
Next post: Generative AI — what it actually is, the risks associated with it, and how it differs from "regular" AI models.