图1 Ground Truth和Predictions
表1 TP和FP统计
Images | Predictions | Confidence | TP or FP |
Images 1 | Person1 | 87% | FP |
Images 1 | Person2 | 80% | FP |
Images 1 | Person3 | 78% | TP |
Images 2 | Person4 | 75% | TP |
Images 2 | Person5 | 52% | FP |
Images 2 | Person6 | 74% | FP |
Images 3 | Person7 | 20% | TP |
Images 3 | Person8 | 68% | FP |
Images 3 | Person9 | 95% | TP |
Images 3 | Person10 | 81% | TP |
Images 4 | Person11 | 90% | TP |
Images 4 | Person12 | 45% | FP |
Images 4 | Person13 | 15% | FP |
Images 5 | Person14 | 98% | TP |
Images 5 | Person15 | 25% | FP |
Images 5 | Person16 | 44% | FP |
Images 6 | Person17 | 43% | FP |
Images 6 | Person18 | 60% | TP |
Images 6 | Person19 | 70% | TP |
Images 6 | Person20 | 84% | FP |
Images 7 | Person21 | 39% | TP |
Images 7 | Person22 | 92% | TP |
Images 8 | Person23 | 36% | TP |
Images 8 | Person24 | 94% | FP |
Images 8 | Person25 | 33% | FP |
表2 P和R统计
Images | Predictions | Confidence | TP | FP | Precision | Recall |
Images 5 | Person14 | 98% | 1 | 0 | 1 | 0.0435 |
Images 3 | Person9 | 95% | 1 | 0 | 1 | 0.0870 |
Images 8 | Person24 | 94% | 0 | 1 | 0.6667 | 0.0870 |
Images 7 | Person22 | 92% | 1 | 0 | 0.7500 | 0.1304 |
Images 4 | Person11 | 90% | 1 | 0 | 0.8000 | 0.1739 |
Images 1 | Person1 | 87% | 0 | 1 | 0.6667 | 0.1739 |
Images 6 | Person20 | 84% | 0 | 1 | 0.5714 | 0.1739 |
Images 3 | Person10 | 81% | 1 | 0 | 0.6250 | 0.2174 |
Images 1 | Person2 | 80% | 0 | 1 | 0.5556 | 0.2174 |
Images 1 | Person3 | 78% | 1 | 0 | 0.6000 | 0.2609 |
Images 2 | Person4 | 75% | 1 | 0 | 0.6364 | 0.3043 |
Images 2 | Person6 | 74% | 0 | 1 | 0.5833 | 0.3043 |
Images 6 | Person19 | 70% | 1 | 0 | 0.6154 | 0.3478 |
Images 3 | Person8 | 68% | 0 | 1 | 0.5714 | 0.3478 |
Images 6 | Person18 | 60% | 1 | 0 | 0.6000 | 0.3913 |
Images 2 | Person5 | 52% | 0 | 1 | 0.5625 | 0.3913 |
Images 4 | Person12 | 45% | 0 | 1 | 0.5294 | 0.3913 |
Images 5 | Person16 | 44% | 0 | 1 | 0.5000 | 0.3913 |
Images 6 | Person17 | 43% | 0 | 1 | 0.4737 | 0.3913 |
Images 7 | Person21 | 39% | 1 | 0 | 0.5000 | 0.4348 |
Images 8 | Person23 | 36% | 1 | 0 | 0.5238 | 0.4783 |
Images 8 | Person25 | 33% | 0 | 1 | 0.5000 | 0.4783 |
Images 5 | Person15 | 25% | 0 | 1 | 0.4783 | 0.4783 |
Images 3 | Person7 | 20% | 1 | 0 | 0.5000 | 0.5217 |
Images 4 | Person13 | 15% | 0 | 1 | 0.4800 | 0.5217 |
通过表2的出了每个Person类别的准确率和召回率,画出PR曲线如图2.18所示。而AP值比较常用的是使用积分法计算。积分法的计算方法很简单,将PR曲线中的各个矩形面积进行相加即可得到AP值,即图2.19中A1 、A2、 A3、 A4、 A5 、A6、 A7的面积相加就是相应的AP值。通过计算可得:
图2.18 PR曲线
图2.19 积分法计算AP