Object Detection
Classification tells you a photo contains a dog. Detection tells you where, how many, and draws a box around each one — a harder problem, and the rest of this part exists to solve it.
Read first: Convolutional Neural Networks
Ask a classifier what is in a photo and it gives you one answer for the entire image, no matter how many things are actually in it. Detection is the harder question underneath: where is each one, and how many are there.
From what, to where
Classification, from a few lessons back, answers exactly one question: given the whole photo, what is the single most likely label. One photo in, one label out, and how many actual objects are in that photo is never asked.
A photo of a street corner might contain three cars, two pedestrians, and a dog. A classifier gives you one word. Detection has to answer a question classification was never built to answer: not just what, but where, and how many — and the count is not known in advance. It could be zero objects or it could be forty.
A box and a label, per object
A detector’s output is concrete, not a vague heat-map of interest. For every object it finds, it returns a tuple: four numbers describing a box (say, the coordinates of its top-left corner plus a width and height), a class label, and a confidence score for that label. A photo with three cars and two pedestrians should come back as five such tuples, not one.
That is the entire contract. A detection pipeline can be built from almost any method underneath, as long as what comes out the other end is a list of (box, label, confidence) — nothing more is asked of it and nothing less will do.
Why one forward pass is not enough
The obvious first idea is to reuse the classifier: run it once, get an answer. That does not work, because a classifier only ever answers “what is the single dominant thing in this entire image”, and a street corner does not have a single dominant thing.
The next idea is to run the classifier again and again, on every possible rectangle the image could contain — every position, every width, every height. A modest 224×224 photo has tens of thousands of plausible box positions at just a handful of sizes; check every size at every position and you are into the hundreds of millions. That is a combinatorial explosion, not a plan, and it is why real detectors take a different route.
Two families of solution
Two real approaches survive contact with that problem, and they trade speed against accuracy in opposite directions.
Two-stage detectors
First propose a modest number of candidate regions likely to contain something, then run a classifier on each candidate. Historically more accurate, because each region gets dedicated attention — and slower, because it is two networks’ worth of work per photo.
Single-stage detectors
Predict every box and class directly in one pass, by dividing the image into a grid and asking each cell what it sees. Faster, often fast enough for real-time video, and the default choice when latency matters more than squeezing out the last point of accuracy.
The chapters ahead build the two ideas a detector actually needs to work at all — a way to say a predicted box is “close enough” (bounding boxes and IoU), and a way to collapse a dozen boxes around one object into one (non-max suppression) — before returning to how either family is graded.
Key takeaways
- Classification answers one question for the whole photo. Detection answers an unknown number of questions — where is each object, and how many are there.
- A detector's real output is a list of tuples: a box, a class label, and a confidence score, one tuple per object found.
- Running a classifier once over an entire photo cannot solve detection, because a classifier only ever names the single dominant thing in front of it.
- Checking every possible box position and size by brute force is a combinatorial explosion, not a plan — a modest photo has millions of plausible boxes.
- Two-stage detectors propose regions before classifying them and tend to be more accurate but slower. Single-stage detectors predict boxes and classes in one pass and tend to be faster.
Before you move on
Quick check
Answer these to unlock the next chapter: 3 of 4 to pass. You can retake it anytime.
Make a free account to read on
Every chapter is free — an account lets your course progress follow you from your laptop to your phone. No payment, no trial.