Table 1 StairNet architecture.
Name | Type | Output size | |||
|---|---|---|---|---|---|
Initial | Tensor slice | 256 \(\times\) 256 \(\times\) 64 | |||
Bottleneck 1.0 | Downsampling | 128 \(\times\) 128 \(\times\) 256 | |||
Bottleneck 1.1 | Â | Â | 128 \(\times\) 128 \(\times\) 256 | ||
Bottleneck 1.2 | Â | Â | 128 \(\times\) 128 \(\times\) 256 | ||
Bottleneck 2.0 | Downsampling | 64 \(\times\) 64 \(\times\) 512 | |||
Bottleneck 2.1 | Dilated (1,2) and (2,2) | 64 \(\times\) 64 \(\times\) 512 | |||
Bottleneck 2.2 | Â | Â | 64 \(\times\) 64 \(\times\) 512 | ||
Bottleneck 2.3 | Dilated (2,4) and (4,4) | 64 \(\times\) 64 \(\times\) 512 | |||
Bottleneck 2.4 | Â | Â | 64 \(\times\) 64 \(\times\) 512 | ||
Bottleneck 2.5 | Dilated (3,8) and (8,8) | 64 \(\times\) 64 \(\times\) 512 | |||
Bottleneck 2.6 | Â | Â | 64 \(\times\) 64 \(\times\) 512 | ||
Bottleneck 2.7 | Dilated (4,16) and (16,16) | 64 \(\times\) 64 \(\times\) 512 | |||
Repeat bottlenecks 2.0 to 2.7 without downsampling operation of bottleneck 2.0 | |||||
ASPP | Â | Â | 64 \(\times\) 64 \(\times\) 512 | ||
Conv 3 \(\times\) 3 | Â | Â | 64 \(\times\) 64 \(\times\) 128 | ||
classification | location | classification | location | classification | location |
Conv 3 \(\times\) 3 | Conv 3 \(\times\) 3 | Â | Â | 64 \(\times\) 64 \(\times\) 128 | 64 \(\times\) 64 \(\times\) 128 |
Conv 1 \(\times\) 1 | Conv 1 \(\times\) 1 | Â | Â | 64 \(\times\) 64 \(\times\) 2 | 64 \(\times\) 64 \(\times\) 8 |
| Â | Sigmoid | Â | Activation | 64 \(\times\) 64 \(\times\) 2 | 64 \(\times\) 64 \(\times\) 8 |
