Intersection over Union (IoU)

Intersection over Union, also known as the Jaccard Index, quantifies the similarity between two regions by computing the area of their overlap divided by the area of their combined extent. IoU ranges from 0 (no overlap) to 1 (perfect overlap). In object detectionObject DetectionObject Detection is a computer vision technique that identifies and localizes specific objects within images or video..., IoU measures how well a predicted bounding boxBounding BoxA bounding box is the minimum axis-aligned rectangle that completely encloses a geographic feature or dataset, define... matches the ground truthGround TruthGround truth refers to data collected at the Earth's surface to validate and calibrate information derived from remot... bounding box. In semantic segmentationSemantic SegmentationSemantic Segmentation is a computer vision technique that assigns a class label to every pixel in an image, enabling ..., IoU is computed per class by comparing predicted and ground truth pixel labels. Mean IoU (mIoU), averaged across all classes, is the most widely used metric for segmentation evaluation. IoU in Geospatial Detection and SegmentationIoU is the standard metric for evaluating geospatial AI tasks that involve spatial extent prediction. Building footprintBuilding FootprintA Building Footprint is the outline of a building as seen from directly above, representing the area of ground it occ... extraction is evaluated by computing IoU between predicted and reference building polygons. Flood extent mapping uses IoU to assess how accurately the model delineates inundated areas. Road network extraction evaluates the overlap between predicted and actual road surfaces. Object detection systems use IoU thresholds to determine whether a detection counts as a true positive: typically, a predicted bounding box with IoU greater than 0.5 against a ground truth box is considered a correct detection. More stringent thresholds like 0.75 evaluate precise localization. Variants and Complementary MetricsSeveral IoU variants address specific evaluation needs. Boundary IoU weights the metric toward boundary pixels, useful for evaluating the sharpness of building or parcel boundaries. Generalized IoU (GIoU) extends IoU to handle non-overlapping predictions by incorporating the smallest enclosing box. Panoptic Quality combines IoU with detection metrics for instance segmentation evaluation. In geospatial applications, IoU is typically reported alongside pixel-level accuracy and F1 scores to provide a comprehensive assessment of model performance across different aspects of prediction quality.