Table 4.
MAP@R: triplet loss semi-hard mining vs. Proxy-NCA on training classes. Bold indicates the best performing method, and grey highlights results that are not statistically significantly different from the best.
| MAP@R | |||||||
|---|---|---|---|---|---|---|---|
| Faces | Flanks | ||||||
| Architecture | Loss | Lions | Chimps | Pandas | Nyala | Zebra | Tiger |
| VGG-11 | Triplet | 16.5 ± 2 | 12.9 ± 2 | 32.0 ± 2 | 11.2 ± 0 | 16.8 ± 1 | 22.8 ± 1 |
| P-NCA | 17.7 ± 1 | 13.8 ± 3 | 31.8 ± 1 | 11.0 ± 1 | 16.5 ± 0 | 22.9 ± 2 | |
| VGG-19 | Triplet | 18.0 ± 2 | 11.7 ± 1 | 25.0 ± 0 | 10.8 ± 1 | 16.7 ± 2 | 21.8 ± 1 |
| P-NCA | 17.7 ± 0 | 12.0 ± 2 | 28.7 ± 0 | 9.7 ± 3 | 16.4 ± 3 | 20.0 ± 1 | |
| ResNet-18 | Triplet | 18.5 ± 0 | 11.2 ± 2 | 26.3 ± 1 | 9.9 ± 2 | 19.0 ± 0 | 24.6 ± 4 |
| P-NCA | 19.0 ± 1 | 11.5 ± 1 | 24.9 ± 0 | 9.5 ± 1 | 18.2 ± 1 | 21.7 ± 2 | |
| ResNet-152 | Triplet | 17.3 ± 2 | 10.1 ± 0 | 26.9 ± 1 | 8.2 ± 0 | 12.1 ± 3 | 12.5 ± 3 |
| P-NCA | 17.1 ± 0 | 9.4 ± 3 | 20.3 ± 1 | 9.0 ± 2 | 11.9 ± 2 | 11.0 ± 1 | |
| DenseNet-201 | Triplet | 20.8 ± 1 | 9.9 ± 2 | 31.1 ± 1 | 11.0 ± 2 | 15.9 ± 2 | 22.3 ± 1 |
| P-NCA | 20.2 ± 2 | 11.6 ± 3 | 28.4 ± 2 | 10.4 ± 1 | 16.0 ± 1 | 23.2 ± 3 | |