Part 3 · Deep Learning & AI

3.4 Object detection and segmentation

Answering “what is where”

10-minute readPrerequisites: 3.3 CNNs and image recognition, 2.7 Evaluation metrics▶ Video (about 5 min)

What you will learn

Explain the difference between image classification, object detection and segmentation, together with their uses in automated driving
Calculate IoU, which measures how well boxes overlap, and explain how NMS removes duplicate boxes
Explain how detection is evaluated (IoU criteria, precision and recall, AP) and the pitfalls with small distant objects, occluded objects and objects side by side

Build intuition

The image classification of 3.3 answers “this image contains a pedestrian”. Driving, however, needs to know what is where, and how many: “one pedestrian ahead on the left, two cars in the right lane”. Several kinds of task serve this purpose:

Task	What it answers	Example in automated driving
Image classification	One category for the whole image	Which sign a cropped sign image shows
Object detection	A box (position and size), category and confidence for each object	Finding pedestrians, cyclists and cars
Semantic segmentation	A category for every pixel	Road, sidewalk, lane markings — where the car can drive
Instance segmentation	A category for every pixel, plus which individual object it belongs to	Separating overlapping pedestrians one by one

Two ways of answering for the same scene (schematic). Object detection gives each object a box, a category and a confidence; segmentation gives every pixel a category.

An object detection model picks up features with a CNN (3.3) or similar, then predicts, all at once for many positions and sizes across the image, “is there an object here, what is it, and how should the box be adjusted?”. This produces many slightly shifted candidate boxes for the same object. So the most confident boxes are kept, duplicates are removed, and only those above the threshold (2.4) become the answer.

Check with math

① IoU: how well a box fits

IoU (introduced in 2.7) expresses how much a predicted box A overlaps the true box B: the area of the overlap divided by the area they cover together.

$$\mathrm{IoU} = \frac{|A \cap B|}{|A \cup B|}$$

∩ is the overlap, ∪ is the combined region, and | | means area (see Notation). A perfect match gives 1; no overlap at all gives 0. Many evaluations count a detection as correct when IoU is at least 0.5.

The smaller the object, the more a slight shift lowers IoU. For a distant pedestrian, a few pixels of error are enough to fall below the “correct detection” criterion.

On the left, the overlap is 80 × 180 = 14,400 and the combined area is 20,000 + 20,000 − 14,400 = 25,600, so IoU ≈ 0.56. On the right, the overlap is 6 × 28 = 168 and the combined area is 280 + 280 − 168 = 392, so IoU ≈ 0.43.

② What a detector outputs, and how it learns

For each object, a detector outputs the box’s position and size (x, y, width, height), a probability for each category, and a confidence. The training loss adds the box error (regression, as in 2.2) to the category error (the cross-entropy of 2.4). In other words, it solves regression and classification at the same time.

③ NMS: removing duplicate boxes

The standard procedure for removing duplicate boxes for the same object is NMS (non-maximum suppression):

Pick the box with the highest confidence and keep it as an answer
Remove boxes whose IoU with that box exceeds a threshold (for example 0.5), treating them as duplicates of the same object
Repeat steps 1–2 with the remaining boxes

This procedure becomes a weakness when boxes of different objects overlap heavily. If the boxes of two pedestrians walking side by side have an IoU of 0.6, NMS with a threshold of 0.5 removes one of them as a “duplicate”.

④ Evaluating detection: AP

For evaluation, predicted boxes are sorted by confidence; those matched to a true box with IoU at or above the criterion count as correct detections, the rest as false alarms, and true boxes with no matching prediction as misses (the confusion matrix of 2.7). Computing precision and recall while moving the threshold gives a curve whose area is the AP (average precision); the mean of the APs across categories is the mAP. For segmentation, the mean IoU across categories (mIoU) is commonly used.

Summary

Object detection gives each object a box, a category and a confidence; segmentation gives every pixel a category. Driving needs “what, where and how many”
Box fit is measured by IoU (overlap ÷ combined area), and duplicate boxes are removed by NMS. Detection is evaluated with precision and recall based on an IoU criterion, and with AP
Small objects can fail the criterion with a few pixels of error, and objects side by side can vanish in NMS. Evaluate separately by distance, path relevance, size and occlusion

Quick quiz

Q1. A true box and a predicted box overlap by 14,400, and together cover 25,600. What is the IoU? With a criterion of 0.5, is it a correct detection?

IoU = 14,400 ÷ 25,600 ≈ 0.56, which is at least 0.5, so it counts as a correct detection.

Q2. For a distant pedestrian 28 pixels tall and 10 wide, the predicted box (the same size) is shifted 4 pixels sideways. What is the IoU?

The overlap is 6 × 28 = 168 and the combined area 280 + 280 − 168 = 392, so IoU ≈ 0.43. With a 0.5 criterion, this prediction counts as a false alarm and the true pedestrian as a miss.

Q3. Two pedestrians walk side by side, and their boxes have an IoU of 0.6. What can happen with an NMS threshold of 0.5?

The less confident box is treated as a “duplicate of the same object” and removed, so it can look as if there is only one person (a miss). Crowded scenes need to be checked, including the NMS threshold.

Q4. What are the names of the task that colors every pixel as “road, sidewalk, lane marking, car”, and the task that also separates overlapping pedestrians one by one?

The first is semantic segmentation; the second is instance segmentation.