Table 4. Comparison with other models on ImageNet dataset(%).
| Modela | Top-1 | Top-5 | Down sampling | From scratcha | Storage (Mbits) | FLOPS (×108) |
|---|---|---|---|---|---|---|
| ResNet-18 (Rastegari et al., 2016) | 69.3% | 89.2% | BConv | Yes | 374.1 | 18.1 |
| XNOR-ResNet-18 (Rastegari et al., 2016) | 51.2% | 69.3% | FPConv | Yes | 33.7 | 1.67 |
| ABC-Net-res18 (Lin, Zhao & Pan, 2017) | 42.7% | 67.6% | BConv | No | 27.5 | 1.59 |
| Bi-Real-Net-18 (Liu et al., 2018) | 56.4% | 79.5% | FPConv | No | 33.6 | 1.63 |
| MoBiNet-k4 (Phan et al., 2020) | 54.4% | 77.5% | BConv | Yes | 36.8 | 0.52 |
| AresB-18 | 54.81 | 78.15 | No conv | Yes | 27.6 | 1.61 |
Note:
a
When the model is trained from scratch, a pretrained model are not used in the weight initialization.