What Is Object Detection?
Learn how object detection combines recognition with location, then compare boxes with image labels and pixel masks.
What Is Object Detection? explained visually
Learn how object detection combines recognition with location, then compare boxes with image labels and pixel masks.
What you will learn
Locate objects, represent boxes, filter predictions, evaluate detections, and compare detector designs.
Combine recognition with location
A photo shows three people waiting beside a bus. You ask a computer what is in the picture, and it answers, “A bus.” That answer is useful, but it leaves out something important. Where is the bus? And where are the people?
Object detection answers both questions. It predicts what certain objects are and where they appear in an image.
For each detected object, the system usually returns a class label, a rectangular bounding box, and a confidence score. The label might be person. The box marks the predicted location. The score expresses how strongly the model supports that detection.
So our street photo could produce four detections: one bus and three people. Each person gets a separate box, even though all three share the same class label.
Compare labels, boxes, and masks
Let’s connect this with what we already know. Image classification assigns a label to an image, or sometimes several labels. It does not normally locate each object. Object detection adds locations for individual objects. Semantic segmentation goes further in a different direction: it assigns a class to each pixel.
A rectangular box will often include some background. A segmentation mask can follow an object’s outline much more closely. And instance segmentation can give separate masks to two objects of the same class.
Choose an output and recognize limitations
These tasks solve different problems. A system counting cars may only need boxes. A system measuring the exact area of a leaf may need a mask.
Detection is also limited by the model and the data used to train it. A detector trained to recognize cars and people will not automatically recognize every possible object. Small objects, shadows, and objects hidden behind other objects can make the task harder.
Connect locations to coordinate representations
Here is the main idea to keep: object detection combines recognition with location. It tells us what an object might be and where to look for it.
Next, we’ll unpack those rectangles and see how a bounding box becomes a few numbers.
Quick check: test your understanding
What two questions does object detection answer?
What an object may be and where it appears.
Does a bounding box follow every object contour?
No. A rectangle commonly includes background; a mask can follow the outline.
Can several detections share a class label?
Yes. Three people can have separate boxes with the same person label.
Accessible referenceLesson 63 transcript321 words · Matches the published narration and captionsOpen transcript
A photo shows three people waiting beside a bus. You ask a computer what is in the picture, and it answers, “A bus.” That answer is useful, but it leaves out something important. Where is the bus? And where are the people?
Object detection answers both questions. It predicts what certain objects are and where they appear in an image.
For each detected object, the system usually returns a class label, a rectangular bounding box, and a confidence score. The label might be person. The box marks the predicted location. The score expresses how strongly the model supports that detection.
So our street photo could produce four detections: one bus and three people. Each person gets a separate box, even though all three share the same class label.
Let’s connect this with what we already know. Image classification assigns a label to an image, or sometimes several labels. It does not normally locate each object. Object detection adds locations for individual objects. Semantic segmentation goes further in a different direction: it assigns a class to each pixel.
A rectangular box will often include some background. A segmentation mask can follow an object’s outline much more closely. And instance segmentation can give separate masks to two objects of the same class.
These tasks solve different problems. A system counting cars may only need boxes. A system measuring the exact area of a leaf may need a mask.
Detection is also limited by the model and the data used to train it. A detector trained to recognize cars and people will not automatically recognize every possible object. Small objects, shadows, and objects hidden behind other objects can make the task harder.
Here is the main idea to keep: object detection combines recognition with location. It tells us what an object might be and where to look for it.
Next, we’ll unpack those rectangles and see how a bounding box becomes a few numbers.