AP50公式检测任务IoU框重叠IoU∣Bpred∩Bgt∣∣Bpred∪Bgt∣ IoU\frac{|B_{pred}\cap B_{gt}|}{|B_{pred}\cup B_{gt}|}IoU∣Bpred​∪Bgt​∣∣Bpred​∩Bgt​∣​在每个类别上这里类别是 Real/Fake按预测分数从高到低排序后逐个匹配贪心若与某个未匹配GT同类且IoU≥0.5IoU\ge 0.5IoU≥0.5 TP_i1否则 FP_i1前缀累计TP(k)∑i1kTPi,FP(k)∑i1kFPi TP(k)\sum_{i1}^{k}TP_i,\quad FP(k)\sum_{i1}^{k} FP_iTP(k)i1∑k​TPi​,FP(k)i1∑k​FPi​精确率/召回率P(k)TP(k)TP(k)FP(k),R(k)TP(k)Ngt P(k)\frac{TP(k)}{TP(k)FP(k)},\quad R(k)\frac{TP(k)}{N_{gt}}P(k)TP(k)FP(k)TP(k)​,R(k)Ngt​TP(k)​插值精确率COCO做法Pinterp(r)max⁡r~≥rP(r~) P_{interp}(r)\max_{\tilde r\ge r}P(\tilde r)Pinterp​(r)r~≥rmax​P(r~)在r∈0,0.01,…,1.00r\in{0,0.01,\dots,1.00}r∈0,0.01,…,1.00101个点上取均值AP50class1101∑rPinterp(r) AP50_{class}\frac{1}{101}\sum_{r}P_{interp}(r)AP50class​1011​r∑​Pinterp​(r)最后对类别求平均忽略无效项得到 AP50。