ML Metrics Detection APmAPAR

[junxnone]

AP & mAP & AR

• mAP@[0.5:.05:.95] - 当前比较流行的衡量标准

AP

• PR 曲线下的面积
• AP 一般指单个类别

计算步骤

GT
Predict Dog

计算 Precision & Recall

• 根据 confidence 排序

• 

Plot PRCurve

• 每个 Recall Level 的 Precision 值 使用右侧最大值 使其单调?

11 Point Interpolation Method

• 11 Recall Point: N=11 [0, 0.1, 0.2, ..., 0,9, 1.0]

• $\huge AP = \frac{1}{11}\times \int_{1}^{11}Precisions$
• $\huge =\frac{1}{11}\times (1+4\times 0.71+6\times 0)$
• $\huge =0.349$

The intention in interpolating the precision/recall curve in this way is to reduce the impact of the “wiggles” in the precision/recall curve, caused by small variations in the ranking of examples.

Interpolating all points

• AUC(Area Under Curve) on PR Curve

$\huge \sum{0}^{n}(r{n+1} - rn)p{interp}(r{n+1})\newline p{interp}(r{n+1})=\underset{\bar{r}:\bar(r)\geq r{n+1}}{max}p(\tilde{r})$

101 Point Interpolation AP

• N=101: MSCOCO AP [0:.01:1]

mAP

• mAP - mAP 一般指所有类别的平均
• $mAP = \frac{\int_{1}^{n} AP}{n}$

---

• mAP@.50IOU - 当 IoU 大于 0.50 时认为正确识别的 mAP
• mAP@.75IOU - 当 IoU 大于 0.75 时认为正确识别时的 mAP
• mAP@[.5:.95] - IoU [0.5, 0.95] 范围内以步长 0.05 间隔计算 mAP

COCO 中称 mAP 为 AP We make no distinction between AP and mAP (and likewise AR and mAR) and assume the difference is clear from context.

AR

• AR - 在每个图像中检测到固定数量的最大召回（recall），在类别和IoU上平均