DistanceIoU

DistanceIoU, commonly abbreviated DIoU, is a bounding box regression loss and evaluation metric used in object detection. It extends the traditional Intersection over Union (IoU) by incorporating the spatial relationship between the predicted box and the ground-truth box, specifically penalizing the distance between their centers in addition to their overlap. This helps improve localization accuracy and training efficiency, especially when boxes have similar IoU but different center positions.

Formally, for a predicted box B and a ground-truth box G, let IoU(B, G) be their intersection over union, ρ(B, G) be the Euclidean distance between their center points, and c be the diagonal length of the smallest enclosing box that covers both B and G. The DistanceIoU loss is defined as L_DIoU = 1 − IoU(B, G) + [ρ(B, G)]^2 / c^2. The first term incentivizes large overlaps, while the second term punishes center misalignment, promoting boxes that are both overlapping and well-centered with respect to the ground-truth.

DistanceIoU is often used as a drop-in replacement for IoU-based regression losses in object detectors to achieve faster convergence and better localization. It is related to, but distinct from, Complete IoU (CIoU), which adds a term accounting for aspect ratio differences between boxes. DIoU is primarily applicable to axis-aligned bounding boxes; for rotated boxes, other metrics are typically employed. Overall, DIoU provides a simple, effective way to improve bounding box regression during model training.