Table 5.
True Positive Rate for related works in comparison to the multi-label, multi-scale IPB configuration, where Multi-scale is denoted as MS, and Multi-label is denoted as ML.
| Config./Class (std.) | Debris | Dense | Diff | Spread | Avg. |
|---|---|---|---|---|---|
| SimCLR25 | 0.8559 (0.0346) | 0.8712 (0.0274) | 0.7538 (0.0592) | 0.7874 (0.0213) | 0.8170 |
| CC27 | 0.5101 (0.1756) | 0.3852 (0.4324) | 0.2277 (0.2370) | 0.3436 (0.0841) | 0.3666 |
| SimTriplet30 | 0.1726 (0.0266) | 0.7160 (0.1390) | 0.2246 (0.1864) | 0.5423 (0.1012) | 0.4138 |
| MPL (Imbalanced)33 | 0.8158 (0.0804) | 0.8423 (0.0407) | 0.8818 (0.0504) | 0.9292 (0.0237) | 0.8672 |
| MPL (Balanced) | 0.8950 (0.0106) | 0.8261 (0.0540) | 0.8322 (0.0564) | 0.9111 (0.0221) | 0.8661 |
| IPB (ML+MS) | 0.8843 (0.0159) | 0.8575 (0.0218) | 0.9085 (0.0189) | 0.8985 (0.0074) | 0.8872 |
Related methods display inferior results due to the effect of class imbalances, which causes over-fitting and high variance.
Bold values represent best result for each column.