Step by Step
C
Classification — what is in this image
Image classification simply answers what is in the image overall, without identifying where in the image it is located.
Example: a classifier simply outputting "dog" as the answer for an entire image, without specifying where in the image the dog actually is.
D
Detection — what AND where
Object detection answers both what is present and where it is located, producing a bounding box plus a class label for each detected object.
Example: a detector outputting a bounding box around a specific dog in the image, along with the label "dog" for that box.
S
Segmentation — what AND which exact pixels
Segmentation goes further still, identifying exactly which pixels belong to each detected object, rather than just a rectangular bounding box.
Example: a segmentation model outlining the exact pixel-level silhouette of a dog, rather than just a rectangular box around it.
I
IoU — measuring detection accuracy
Intersection over Union (IoU) measures the overlap between a predicted bounding box and the ground truth box, calculated as the area of overlap divided by the area of union — a threshold of 0.5 is the standard cutoff for an acceptable detection.
Example: a predicted bounding box overlapping 70% with the ground truth box producing an IoU of 0.7, comfortably above the standard 0.5 threshold for acceptable detection.
Applied Walkthrough
1
A system needs to identify not just that a photo contains a dog, but exactly where the dog is located and its precise pixel outline.
2
Ask: which single task — classification, detection, or segmentation — accomplishes all of this at once? None alone; classification only answers "what," detection adds "where" via a bounding box, and only segmentation adds the exact pixel-level "which pixels" answer.
3
A detector like YOLO (You Only Look Once) can quickly produce the bounding box (the "where"), evaluated using IoU to measure how well that predicted box overlaps with the true ground truth box.
4
If a more precise, pixel-level outline is specifically needed rather than just a bounding box, a segmentation model would be required instead of (or in addition to) the detection step.
Exam Application
Exams test whether you can correctly distinguish classification (what), detection (what + where via bounding box), and segmentation (what + which exact pixels), and whether you can calculate or interpret IoU values against the standard 0.5 acceptable-detection threshold.
⚠ Common Trap
The most common trap is confusing detection with segmentation, since both involve locating objects. Detection produces only a rectangular bounding box (an approximate location); segmentation identifies the exact pixels belonging to the object — a meaningfully more precise and computationally demanding task.
✓ Quick Self-Check
1. What does image classification answer?
What is in the image, without specifying location.
Tap to reveal / hide
2. What does object detection answer, beyond classification?
What AND where — producing a bounding box plus class label for each object.
Tap to reveal / hide
3. What does segmentation answer, beyond detection?
What AND which exact pixels belong to each object, rather than just a bounding box.
Tap to reveal / hide
4. What does IoU measure, and how is it calculated?
The overlap between a predicted bounding box and the ground truth box, calculated as area of overlap divided by area of union.
Tap to reveal / hide
5. What is the standard IoU threshold for an acceptable detection?
0.5.
Tap to reveal / hide