Intersection over Union - IoU - 交并比 (一)
Intersection over Union - IoU - 交并比 (一)
1. Intersection over UnionThe IoU is the ratio of the overlapping area of ground truth and predicted area to the total area. Here is a visual explanation of the metric:
The two squares represent the bounding boxes of ground truth and predictions. The IoU is calculated as a ratio of the area of overlap to the area of the union. Here is the script to compute the IoU is given ground truth and prediction bounding boxes.
def calculate_iou(gt_bb, pred_bb)
gt_bb: ground truth bounding box
pred_bb: prediction bounding box
2. Intersection over Union (IoU，交并比)
目标检测的评价体系中，有一个参数叫做 IoU (交并比)，即模型产生的目标窗口与原来标记窗口的交叠率。检测结果 (Detection Result) 与 Ground Truth 的交集除以它们的并集，即为检测的准确率 IoU。
理想情况下，IOU = 1，即完全重合。
GT = GroundTruth
DR = DetectionResult
交集：GT ⋂ DR
并集：GT ⋃ DR
蓝色的框：GroundTruth
黄色的框：DetectionResult
绿色的框：DetectionResult ⋂ GroundTruth
红色的框：DetectionResult ⋃ GroundTruth
最理想的情况就是 DR 与 GT 完全重合，即 IOU = 1。
评价一个算法的时候，一种常见的方法是先设置一个 IOU 的阈值，只要算法找到的 IOU 大于这个阈值，就是一个有效的检测。一般来说，这个 score > 0.5 就可以被认为一个不错的结果。