Table 3.
Evaluation of iNat2021 mini dataset on segmentation medical tasks. Even with less than half number of pre-training samples, iNat2021 mini achieves equal or superior performance over ImageNet counterpart. Best performance is bolded and second best is underlined.
Pre-training task | Target tasks | |||
---|---|---|---|---|
Dataset | #training data | PXS | VFS | LXS |
ImageNet | 1.3M | 67.93±1.45 | 79.20±0.34 | 98.19±0.13 |
iNat2021 mini | 500K | 68.26±1.48 | 79.24±0.28 | 98.19±0.09 |
iNat2021 | 2.7M | 68.43±0.92 | 79.33±0.18 | 98.25±0.07 |