3.2.5. Hyperparameters Tuning

We tried different semantic segmentation network architectures. For SegNet and DeepLab v3+ we tuned hyperparameters according to Table 4. Please note that DeepLab v3+ with ResNet-18 backbone is more lightweight than SegNet, and this allowed us to use a larger mini-batch size, with eight patches per mini-batch. With our GPU, SegNet was trained with only one patch per mini-batch. More details about hyperparameters can be found in MATLAB documentation [33].


