Open junxnone opened 1 year ago
mAP@[0.5:.05:.95]
confidence
Recall Level
Precision
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.
$\huge \sum{0}^{n}(r{n+1} - rn)p{interp}(r{n+1})\newline p{interp}(r{n+1})=\underset{\bar{r}:\bar(r)\geq r{n+1}}{max}p(\tilde{r})$
[0:.01:1]
mAP@.50IOU
mAP@.75IOU
mAP@[.5:.95]
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.
AP & mAP & AR
mAP@[0.5:.05:.95]
- 当前比较流行的衡量标准AP
计算步骤
计算 Precision & Recall
confidence
排序Plot PRCurve
Recall Level
的Precision
值 使用右侧最大值 使其单调?11 Point Interpolation Method
Interpolating all points
$\huge \sum{0}^{n}(r{n+1} - rn)p{interp}(r{n+1})\newline p{interp}(r{n+1})=\underset{\bar{r}:\bar(r)\geq r{n+1}}{max}p(\tilde{r})$
101 Point Interpolation AP
[0:.01:1]
mAP
mAP@.50IOU
- 当 IoU 大于 0.50 时认为正确识别的 mAPmAP@.75IOU
- 当 IoU 大于 0.75 时认为正确识别时的 mAPmAP@[.5:.95]
- IoU [0.5, 0.95] 范围内以步长 0.05 间隔计算 mAPAR