A retail analytics team wants a vision model that both draws a box around each product on a shelf image and labels what product it is. Which computer-vision task does that describe?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Here's the deal: object detection is "where is it and what is it?" Your boss drops a shelf photo and says, "Count every SKU and tell me which ones." You don't just say "grocery aisle" — you put boxes on each item and name them. Think of it like a security guard who points at people and says who they are, not someone who only says "busy lobby." Exam trap: mixing detection with classification (whole image), segmentation (every pixel), or generative image creation. Those are different jobs. For the test: boxes + labels = object detection. Got it? Sweet — lock that in.
Full explanation below image
Full Explanation
Object detection is a core computer-vision task that jointly localizes objects in an image or video frame and classifies each localized instance. Typical outputs are bounding boxes (or sometimes oriented boxes) plus class labels and confidence scores. Multi-object scenes—shelves, traffic, manufacturing lines—are the natural setting for detection because a single global label is too coarse for inventory, safety, or analytics use cases that need per-item presence and position.
The correct answer matches that definition: locate objects and assign classes. Modern detectors include two-stage families (region proposal then classify/refine) and single-stage families (predict boxes and classes in one pass). Evaluation often uses precision and recall at intersection-over-union (IoU) thresholds and mean Average Precision (mAP). In production, teams also track latency, false positives on background clutter, and failure modes under occlusion or unusual lighting.
Text-to-image synthesis is generative: text goes in, a new image comes out. That is the opposite workflow from analyzing an existing shelf photo; it creates pixels rather than reading them. Semantic segmentation paints every pixel with a class; it is denser than detection and usually does not output discrete instance boxes (instance segmentation is related but distinct). Image classification produces one or a few labels for the whole image and provides no spatial localization of individual products—so “grocery aisle” is not the same as “twelve SKUs with boxes.”
Underlying principle: task choice follows the required output structure. If the business needs both where and what for multiple instances, detection is the right abstraction. Best practice is to match annotation guidelines to that output (box quality, class taxonomy) and to evaluate at realistic IoU and confidence thresholds. Memory aid: classification equals “what overall”; detection equals “what and where (instances)”; segmentation equals “which class for each pixel.” When a requirement mentions both position and identity of multiple objects, choose object detection.