CIoU损失函数是谁提出的

CIoU（Complete Intersection over Union）损失函数是由 Zheng et al. 在其论文《Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression》中提出的[^3]。该团队通过分析 IoU 和 GIoU 的不足之处，进一步引入了距离因素以及宽高比例因子来提升边界框回归的效果。 以下是 CIoU 损失函数的核心概念及其优势： - CIoU 不仅考虑了预测框与真实框之间的重叠面积（即 IoU），还加入了中心点的距离惩罚项和宽高比例一致性约束[^3]。 - 这种设计使得 CIoU 能够更全面地衡量两个框之间的差异，从而加速收敛并提高定位精度。 ```python import torch def ciou_loss(pred_boxes, target_boxes, eps=1e-7): """ Compute the CIoU loss between predicted boxes and target boxes. Args: pred_boxes (Tensor): Predicted bounding boxes with shape [N, 4]. target_boxes (Tensor): Target bounding boxes with shape [N, 4]. eps (float): Epsilon to avoid division by zero. Returns: Tensor: The computed CIoU loss value. """ # Convert center-size format to top-left bottom-right coordinates px1y1 = pred_boxes[:, :2] - pred_boxes[:, 2:] / 2 px2y2 = pred_boxes[:, :2] + pred_boxes[:, 2:] / 2 tx1y1 = target_boxes[:, :2] - target_boxes[:, 2:] / 2 tx2y2 = target_boxes[:, :2] + target_boxes[:, 2:] / 2 # Calculate intersection area inter_x1y1 = torch.max(px1y1, tx1y1) inter_x2y2 = torch.min(px2y2, tx2y2) inter_area = torch.clamp((inter_x2y2 - inter_x1y1), min=0).prod(1) # Calculate union area pred_area = pred_boxes[:, 2:].prod(1) target_area = target_boxes[:, 2:].prod(1) union_area = pred_area + target_area - inter_area + eps iou = inter_area / union_area # Enclosing box dimensions enclose_x1y1 = torch.min(px1y1, tx1y1) enclose_x2y2 = torch.max(px2y2, tx2y2) enclose_diagonal_sq = ((enclose_x2y2 - enclose_x1y1)**2).sum(dim=1) # Center point distance squared centers_distance_sq = ((pred_boxes[:, :2] - target_boxes[:, :2])**2).sum(dim=1) diou_term = centers_distance_sq / (enclose_diagonal_sq + eps) v = (4 / math.pi ** 2) * torch.pow( torch.atan(target_boxes[:, 2] / (target_boxes[:, 3] + eps)) - torch.atan(pred_boxes[:, 2] / (pred_boxes[:, 3] + eps)), 2 ) alpha = v / (1 - iou + v + eps) ciou = iou - diou_term - alpha * v return 1 - ciou.mean() ```