Detection & Segmentation

IoU ranges from 0 (no overlap) to 1 (perfect match). Work a concrete case. Let the predicted box be the square $(0, 0, 100, 100)$ and the ground-truth box be $(10, 10, 110, 110)$, same $100 \times 100$ size, each offset by 10 pixels.

• Each box has area $100 \times 100 = 10{,}000$.
• The overlap is the square from $(10, 10)$ to $(100, 100)$: $90 \times 90 = 8{,}100$.
• The union is $10{,}000 + 10{,}000 - 8{,}100 = 11{,}900$.
• Therefore $\text{IoU} = 8{,}100 / 11{,}900 \approx 0.68$.

Detectors count a prediction as correct only when its IoU with a ground-truth box clears a threshold. At the common threshold of $0.5$, this prediction passes; a stricter threshold of $0.75$ would reject the very same box, since $0.68$ falls below it. That sensitivity to the threshold is exactly why modern benchmarks average over many of them.

From boxes to detections: non-max suppression

A detector does not emit one clean box per object. It proposes many overlapping boxes around each object, each with a confidence score. Non-max suppression (NMS) prunes this thicket down to one box per object.

1. 1Sort by confidenceOrder all candidate boxes for a class from highest predicted score to lowest.
2. 2Take the top boxMove the highest-scoring remaining box to the final list of detections; it is the best guess for one object.
3. 3Suppress its neighborsRemove every other box whose IoU with that box exceeds a threshold such as 0.5, since they describe the same object.
4. 4RepeatReturn to the highest-scoring box still remaining and continue until no candidates are left.

Set the NMS IoU threshold too low and you delete genuine nearby objects; too high and you keep duplicate boxes. It is a real hyperparameter, not an afterthought.

Measuring a detector: mean Average Precision

A single accuracy number cannot capture detection quality, because a detector trades off finding more objects (recall) against being right when it fires (precision), and that trade-off shifts as you move the confidence cutoff. The field's answer is mean Average Precision (mAP).

One-stage versus two-stage detectors

Detector architectures split into two families, trading speed against accuracy, a gap that has narrowed sharply over time.

Segmentation: labeling every pixel

Segmentation pushes localization to its limit, a per-pixel decision. Semantic segmentation outputs a map the same height and width as the input, with a class id at every pixel; architecturally this is usually an encoder that downsamples to extract features followed by a decoder that upsamples back to full resolution (a U-Net-style design). Instance segmentation adds a mask per detected object: Mask R-CNN, for example, extends a two-stage detector with a small branch that predicts a binary foreground mask inside each box. IoU reappears here too, now measured between predicted and true masks rather than boxes, as the natural overlap metric for pixel regions.

Further reading

• COCO detection evaluation, the precise definition of AP, mAP, and the IoU-threshold averaging used as the field's standard.
• Faster R-CNN (Ren et al., 2015), the region proposal network and the canonical two-stage detector.
• You Only Look Once: YOLO (Redmon et al., 2016), the one-stage, real-time detection approach.
• Mask R-CNN (He et al., 2017), extending detection with a per-instance mask head for instance segmentation.