# Human 測試 similarity - 內建

## 使用 ROC

接下來套用模板, 混淆矩陣

### model - threshold

| 真同個人 | 偽同個人 |
| -------- | ------ |
| 偽不同人 | 真不同人 |

* TPR:
* FPR:
* ACC:
* ave ACC:
PS: (真同個人 / 實際同個人 + 真不同人 / 實際不同人) / 2
* 除了列出 0.4 - 0.9
會嘗試找促使 ACC, ave ACC 較高的 threshold

---

### HSE FaceRes(Deep) - 0.4

| 1210 | 3596 |
| -------- | ------ |
| 0 | 4 |

* TPR: 100 %
* FPR: 99.89 %
* ACC: 25.24 %

### HSE FaceRes(Deep) - 0.5

| 1209 | 3475 |
| -------- | ------ |
| 1 | 125 |

* TPR: 99.92 %
* FPR: 96.53 %
* ACC: 27.73 %

### HSE FaceRes(Deep) - 0.6

| 1163 | 1866 |
| -------- | ------ |
| 47 | 1734 |

* TPR: 96.12 %
* FPR: 51.83 %
* ACC: 60.23 %
* ave ACC: 72.14 %

### HSE FaceRes(Deep) - 0.64

| 1041 | 806 |
| -------- | ------ |
| 169 | 2794 |

* TPR: 86.03 %
* FPR: 22.39 %
* ACC: 79.73 %
* ave ACC: 81.82 %

### HSE FaceRes(Deep) - 0.65

| 984 | 590 |
| -------- | ------ |
| 226 | 3010 |

* TPR: 81.32 %
* FPR: 16.39 %
* ACC: 83.04 %
* ave ACC: 82.47 %

### HSE FaceRes(Deep) - 0.67

| 984 | 252 |
| -------- | ------ |
| 376 | 3348 |

* TPR: 68.93 %
* FPR: 7.00 %
* ACC: 86.94 %
* ave ACC: 80.96 %

### HSE FaceRes(Deep) - 0.68

| 752 | 138 |
| -------- | ------ |
| 458 | 3462 |

* TPR: 62.15 %
* FPR: 3.83 %
* ACC: 87.61 %
* ave ACC: 79.16 %

### HSE FaceRes(Deep) - 0.69

| 657 | 76 |
| -------- | ------ |
| 553 | 3524 |

* TPR: 54.30 %
* FPR: 2.11 %
* ACC: 86.92 %

### HSE FaceRes(Deep) - 0.7

| 553 | 38 |
| -------- | ------ |
| 657 | 3562 |

* TPR: 45.70 %
* FPR: 1.06 %
* ACC: 85.55 %

### HSE FaceRes(Deep) - 0.8

| 39 | 0 |
| -------- | ------ |
| 1171 | 3600 |

* TPR: 3.22 %
* FPR: 0 %
* ACC: 75.65 %

### HSE FaceRes(Deep) - 0.9

| 3 | 0 |
| -------- | ------ |
| 1207 | 3600 |

* TPR: 0.25 %
* FPR: 0 %
* ACC: 74.91 %

### ROC 圖(點旁的標註為 threshold)

藍色: highest ave ACC
綠色: highest ACC


ave detect single image: 102.58 ms

---

### HSE FaceRes - 0.4

| 1190 | 1467 |
| -------- | ------ |
| 20 | 2133 |

* TPR: 98.35 %
* FPR: 40.75 %
* ACC: 69.09 %
* ave ACC: 78.80 %

### HSE FaceRes - 0.44

| 1128 | 549 |
| -------- | ------ |
| 82 | 3051 |

* TPR: 93.22 %
* FPR: 15.25 %
* ACC: 86.88 %
* ave ACC: 88.99 %

### HSE FaceRes - 0.45

| 1090 | 403 |
| -------- | ------ |
| 120 | 3197 |

* TPR: 90.08 %
* FPR: 11.19 %
* ACC: 89.13 %
* ave ACC: 89.44 %

### HSE FaceRes - 0.47

| 987 | 220 |
| -------- | ------ |
| 223 | 3380 |

* TPR: 81.57 %
* FPR: 6.11 %
* ACC: 90.79 %
* ave ACC: 87.73 %

### HSE FaceRes - 0.48

| 924 | 135 |
| -------- | ------ |
| 286 | 3465 |

* TPR: 76.36 %
* FPR: 3.75 %
* ACC: 91.25 %
* ave ACC: 86.31 %

### HSE FaceRes - 0.49

| 862 | 91 |
| -------- | ------ |
| 348 | 3509 |

* TPR: 71.24 %
* FPR: 2.53 %
* ACC: 90.87 %
* ave ACC: 84.36 %

### HSE FaceRes - 0.5

| 795 | 63 |
| -------- | ------ |
| 415 | 3537 |

* TPR: 65.70 %
* FPR: 1.75 %
* ACC: 90.06 %
* ave ACC: 81.98 %

### HSE FaceRes - 0.6

| 208 | 0 |
| -------- | ------ |
| 1002 | 3600 |

* TPR: 17.19 %
* FPR: 0 %
* ACC: 79.17 %
* ave ACC: 58.60 %

### HSE FaceRes - 0.7

| 41 | 0 |
| -------- | ------ |
| 1169 | 3600 |

* TPR: 3.42 %
* FPR: 0 %
* ACC: 75.70 %

### HSE FaceRes - 0.8

| 4 | 0 |
| -------- | ------ |
| 1206 | 3600 |

* TPR: 0.33 %
* FPR: 0 %
* ACC: 74.93 %

### HSE FaceRes - 0.9

| 2 | 0 |
| -------- | ------ |
| 1208 | 3600 |

* TPR: 0.17 %
* FPR: 0 %
* ACC: 74.89 %

### ROC 圖(點旁的標註為 threshold)

藍色: highest ave ACC
綠色: highest ACC


ave detect single image: 104.17 ms

---

### mobileface - 0.4/0.5/0.6/0.8/0.9

| 1210 | 3600 |
| -------- | ------ |
| 0 | 0 |

* TPR: 100 %
* FPR: 100 %
* ACC: 25.16 %

### mobileface - 0.95

| 1141 | 3326 |
| -------- | ------ |
| 69 | 274 |

* TPR: 94.30 %
* FPR: 92.39 %
* ACC: 29.42 %
* ave ACC: 50.95 %

### mobileface - 0.96

| 891 | 2249 |
| -------- | ------ |
| 319 | 1351 |

* TPR: 73.64 %
* FPR: 62.47 %
* ACC: 46.61 %
* ave ACC: 55.58 %

### mobileface - 0.97

| 434 | 666 |
| -------- | ------ |
| 776 | 2934 |

* TPR: 35.87 %
* FPR: 18.50 %
* ACC: 70.02 %
* ave ACC: 58.68 %

### mobileface - 0.98

| 65 | 27 |
| -------- | ------ |
| 1145 | 3573 |

* TPR: 5.37 %
* FPR: 0.75 %
* ACC: 75.63 %
* ave ACC: 52.31 %

### mobileface - 0.99

| 0 | 0 |
| -------- | ------ |
| 1210 | 3600 |

* TPR: 0 %
* FPR: 0 %
* ACC: 74.84 %
* ave ACC: 50.00 %

### ROC 圖(點旁的標註為 threshold)

藍色: highest ave ACC
綠色: highest ACC


ave detect single image: 82.40 ms

---

### mobilefacenet - 0.4/0.5/0.6/0.8/0.9

| 1210 | 3600 |
| -------- | ------ |
| 0 | 0 |

* TPR: 100 %
* FPR: 100 %
* ACC: 25.16 %

### mobilefacenet - 0.998

| 1210 | 3487 |
| -------- | ------ |
| 0 | 113 |

* TPR: 100 %
* FPR: 96.86 %
* ACC: 27.51 %
* ave ACC: 51.57 %

### mobilefacenet - 0.9985

| 1187 | 2983 |
| -------- | ------ |
| 23 | 617 |

* TPR: 98.10 %
* FPR: 82.86 %
* ACC: 37.51 %
* ave ACC: 57.62 %

### mobilefacenet - 0.999

| 941 | 1663 |
| -------- | ------ |
| 269 | 1937 |

* TPR: 77.77 %
* FPR: 46.19 %
* ACC: 59.83 %
* ave ACC: 65.79 %

### mobilefacenet - 0.9993

| 534 | 660 |
| -------- | ------ |
| 676 | 2940 |

* TPR: 44.13 %
* FPR: 18.33 %
* ACC: 59.83 %
* ave ACC: 62.90 %

### mobilefacenet - 0.9995

| 251 | 201 |
| -------- | ------ |
| 959 | 3399 |

* TPR: 20.74 %
* FPR: 5.58 %
* ACC: 75.88 %
* ace ACC: 57.58 %

### mobilefacenet - 0.9997

| 43 | 28 |
| -------- | ------ |
| 1167 | 3572 |

* TPR: 3.55 %
* FPR: 0.78 %
* ACC: 75.16 %
* ace ACC: 51.39 %

### ROC 圖(點旁的標註為 threshold)

藍色: highest ave ACC
綠色: highest ACC


ave detect single image: 87.03