Table 2. Experimental results of the APFP method proposed in this article on the CUB dataset are presented.
Bold indicates the best experimental results, while italic indicates the second best. § denotes results reproduced using the settings outlined in this article. * taken from Tang, Huang & Zhang (2020).
| Method | Backbone | 5-way 1-shot | 5-way 5-shot |
|---|---|---|---|
| ProtoNet (Snell, Swersky & Zemel, 2017) | Conv4 | 64.42 0.48 | 81.82 0.35 |
| FEAT (Ye et al., 2020) | Conv4 | 68.87 0.22 | 82.90 0.15 |
| MELR (Fei et al., 2021) | Conv4 | 70.26 0.50 | 85.01 0.32 |
| WPN (Zhou & Yu, 2023) | Conv4 | – | 87.03 0.65 |
| SetFeat (Afrasiyabi et al., 2022) | SF-12 | 79.60 0.80 | 90.48 0.44 |
| MatchNet (Vinyals et al., 2016) | ResNet-12 | 71.87 0.85 | 85.08 0.57 |
| BD-CSPN (Liu, Song & Qin, 2020) | ResNet-12 | 84.90 | 90.22 |
| MLCN (Dang et al., 2023) | ResNet-12 | 77.96 0.44 | 91.20 0.24 |
| AA (Afrasiyabi, Lalonde & Gagné, 2020) | ResNet-18 | 74.22 1.09 | 88.65 0.55 |
| ProtoNet§ (Snell, Swersky & Zemel, 2017) | ResNet-18 | 80.85 0.43 | 89.95 0.23 |
| MAML* (Finn, Abbeel & Levine, 2017) | ResNet-18 | 68.42 1.07 | 83.47 0.62 |
| Neg-Cosine (Liu et al., 2020) | ResNet-18 | 72.66 0.85 | 89.40 0.43 |
| LaplacianShot (Ziko et al., 2020) | ResNet-18 | 80.96 | 88.68 |
| Baseline++ (Chen et al., 2019) | ResNet-18 | 67.02 0.90 | 83.58 0.54 |
| FRN (Wertheimer, Tang & Hariharan, 2021) | ResNet-18 | 82.55 0.19 | 92.98 0.10 |
| AAP2S (Ma et al., 2022) | ResNet-18 | 77.64 0.19 | 90.43 0.18 |
| Meta-DeepBDC (Xie et al., 2022) | ResNet-18 | 83.55 0.40 | 93.82 0.17 |
| APFP (ours) | ResNet-18 | 84.02 0.40 | 94.44 0.17 |