The setting of decoder

[SY-Xuan]

Could you please explain that why you reduce the channel of low level features to 48 in the decoder? It is confusing.

BriNet/decoder.py

Line 10 in 82ebe73

self.conv1 = nn.Conv2d(low_level_inplanes, 48, 1, bias=False)

BriNet/decoder.py

Line 11 in 82ebe73

self.bn1 = BatchNorm(48)

BriNet/decoder.py

Line 12 in 82ebe73

self.relu = nn.ReLU()

BriNet/decoder.py

Line 13 in 82ebe73

#channels should be 304

[Wi-sc]

Thanks for your interests. It's just consistent with Deeplab V3+. I think they reduce the channel for balance between memory and effect. You can find the ablation study (Table 1) in their paper.

[SY-Xuan]

Thank you！I understand now.