GeorgeSeif / Semantic-Segmentation-Suite

Semantic Segmentation Suite in TensorFlow. Implement, train, and test new Semantic Segmentation models easily!
2.5k stars 880 forks source link

AdapNet implementation #224

Open kame-hameha opened 4 years ago

kame-hameha commented 4 years ago

Hey Georg, I have do not fully comprehend your AdapNet implementation, specifically lines 119 through 124: 119 net = ResNetBlock_1(net, filters_1=128, filters_2=512)

121 skip_connection = ConvBlock(net, n_filters=12, kernel_size=[1, 1])

124 net = MultiscaleBlock_1(net, filters_1=128, filters_2=512, filters_3=64, p=1, d=2)

According to the paper, the first MultiscaleBlock_1 should be added after the six "vanilla" ResNet blocks, and before doing the ConvBlock with kernel_size=[1, 1] in line 121. I guess the dimensions that are passed through via the skip connection would be somehow different in that case. Or am I misunderstanding something here?