IoU (Intersection over Union)
Intersection over Union (IoU) is a metric in computer vision used to evaluate object detection and segmentation accuracy. It is calculated as the area of overlap divided by the area of union between a predicted bounding box and the ground truth.
Background
IoU has become the standard evaluation metric in object detection benchmarks like PASCAL VOC and COCO. A common threshold (e.g., IoU ≥ 0.5) is used to classify predictions as correct detections. Higher thresholds (e.g., 0.75 or 0.9) are used in applications requiring stricter precision, such as medical imaging.
Examples
- Autonomous driving: detecting pedestrians, vehicles, and traffic signs.
- Medical imaging: assessing tumor segmentation quality.
- Robotics: ensuring reliable detection of objects in manipulation tasks.
Strengths and challenges
- ✅ Simple, widely recognized, and easy to compute.
- ✅ Standard metric in AI competitions.
- ❌ Binary thresholding may penalize near-correct detections.
- ❌ Less effective for evaluating small or overlapping objects.
IoU is popular because of its simplicity and interpretability: a score of 1.0 means perfect overlap, while 0.0 indicates no intersection at all. This makes it intuitive for researchers, engineers, and even non-technical stakeholders to understand model performance.
In practice, IoU is not only used in academic benchmarks but also in production pipelines. For example, quality assurance teams in autonomous driving projects often rely on IoU distributions across large datasets to validate that object detection remains consistent under varying conditions such as rain, night driving, or camera distortion.
That said, IoU has well-known weaknesses. A slightly shifted bounding box can lead to a sharp drop in IoU, even if the prediction is visually acceptable. To address this, alternative metrics have been proposed, such as Generalized IoU (GIoU), Distance IoU (DIoU), and Complete IoU (CIoU), which consider distance, shape, and size alignment in addition to overlap. These refinements make evaluations more robust in real-world scenarios where bounding boxes are rarely perfect.
📚 Further Reading
- Lin, T.-Y. et al. (2014). Microsoft COCO: Common Objects in Context.
- IoU in AI: Key to Precision in Computer Vision, Innovatiana
