Object detection is a type of artificial intelligence that finds and labels objects in images or video. Instead of only saying what is in a picture, it also shows where each object is by drawing a box around it. This matters for tools like self-driving cars, phone cameras, wildlife monitoring, sports analysis, and assistive technology.
It connects computer science with statistics because the AI must make predictions and estimate how confident it is.
Key Facts
- Object detection answers two questions: What object is present and where is it located?
- A bounding box is often written as (x, y, width, height) or as corner points (x1, y1, x2, y2).
- Confidence score is a probability-like value, such as 0.92, that estimates how sure the model is about a detection.
- A model can detect multiple objects in one image, such as person, dog, bicycle, backpack, stop sign, and soccer ball.
- Intersection over Union measures box overlap: IoU = area of overlap / area of union.
- Training improves the model by reducing error between predicted boxes and labels and the correct boxes and labels.
Vocabulary
- Object Detection
- Object detection is an AI task that identifies objects in an image and locates each one with a bounding box.
- Bounding Box
- A bounding box is a rectangle drawn around an object to show its predicted position in an image.
- Confidence Score
- A confidence score is a number that shows how certain the model is about a predicted object label.
- Training Data
- Training data is a collection of labeled examples that a machine learning model studies to learn patterns.
- Intersection over Union
- Intersection over Union is a measure of how much a predicted bounding box overlaps the correct bounding box.
Common Mistakes to Avoid
- Confusing classification with detection. Classification labels the whole image, while detection labels and locates each object inside the image.
- Treating confidence scores as guaranteed truth. A score of 0.90 means the model is very confident, but it can still be wrong if the image is unclear or the training data is biased.
- Ignoring false positives and false negatives. A false positive labels something that is not there, while a false negative misses an object that is there.
- Assuming bigger training data is always better. Training data must be accurate, varied, and well labeled, or the model may learn unreliable patterns.
Practice Questions
- 1 An object detector finds a dog with confidence 0.87, a bicycle with confidence 0.64, and a soccer ball with confidence 0.41. If the display threshold is 0.60, which detections will be shown?
- 2 A predicted box and the correct box overlap by 30 square pixels. The total area covered by both boxes together is 50 square pixels. Calculate IoU = area of overlap / area of union.
- 3 A model detects backpacks very well in classroom photos but often misses them in outdoor sports photos. Explain one likely reason and one way to improve the model.