DIoU

DIoU, short for Distance-IoU, is a bounding-box regression loss used in object detection to improve localization accuracy. It extends the traditional IoU-based loss by incorporating the distance between the centers of the predicted and ground-truth boxes into the objective, encouraging faster and more precise alignment of boxes even when there is substantial overlap.

Mathematically, let Bp be the predicted axis-aligned box with center (xp, yp) and ground-truth box Bg have center (xg, yg). Let IoU denote the intersection-over-union between Bp and Bg. Let d^2 be the squared Euclidean distance between the centers, and let c^2 be the squared diagonal length of the smallest enclosing box that contains both Bp and Bg. Then the DIoU value is defined as DIoU = IoU − d^2 / c^2, and the corresponding loss used for training is L_DIoU = 1 − DIoU = 1 − IoU + d^2 / c^2. This formulation adds a center-distance penalty to the IoU objective, pushing predicted boxes toward the ground-truth centers.

DIoU is differentiable and can be implemented as a drop-in replacement for IoU-based regression losses in many detectors, including anchor-based and some anchor-free architectures. It is typically used alongside a classification loss in fully supervised object detection pipelines. Variants like CIOU (Complete IoU) extend DIoU by incorporating aspect-ratio information to further improve regression accuracy. DIoU is valued for improving convergence speed and localization precision, particularly when boxes overlap but are offset.

---