The following table describes training hyperparameters.

• batch normalization statistics computed per batch during training and using a rolling average computed over input batches for validation and testing
  • original paper uses a static statistics computed over the training data
• encoder transfer learning from VGG16 trained on ImageNet
• best model in terms of validation accuracy is kept as final model
• median frequency balancing of class labels (Eigen et al. 2014)
  • weighted categorical cross-entropy loss function
• local contrast normalization of inputs (Jarrett et al. 2009)
• pooling indexes (Badrinarayanan et al. 2015)

The following table outlines the testing results from SegNet.

The following table describes training hyperparameters.

• batch normalization statistics computed per batch during training and using a rolling average computed over input batches for validation and testing
  • original paper uses a static statistics computed over the training data
• encoder transfer learning from VGG16 trained on ImageNet
  • note that VGG16 does not have any dropout by default; transfer from a Bayesian VGG16 model could improve results
• best model in terms of validation accuracy is kept as final model
• median frequency balancing of class labels (Eigen et al. 2014)
  • weighted categorical cross-entropy loss function
• local contrast normalization of inputs (Jarrett et al. 2009)
• pooling indexes (Badrinarayanan et al. 2015)

The following table outlines the testing results from Bayesian SegNet.

The following table describes training hyperparameters.

• random horizontal flips of images during training
  • the paper says vertical, but their implementation clearly shows horizontal flips (likely a typo). Horizontal make more sense than vertical anyway and produces empirically better test results
• batch normalization statistics computed per batch during training, validation, and testing
• skip connections between encoder and decoder (Jégou et al. 2016)

The following table outlines the testing results from 103 Layers Tiramisu.

The following table describes training hyperparameters.

• network split to predict targets and loss attenuation
  • custom loss function to train the second head of the network (Kendall et al. 2017)
  • our loss function samples through the Softmax function like their paper says (but contrary to the mathematics they present?). without applying the Softmax function, the loss is unstable and goes negative
• pre-trained with fine weights from original Tiramisu
• pre-trained network frozen while head to predict sigma is trained

The quantitative results are the same as the standard Tiramisu model.

• pre-trained with fine weights from original Tiramisu
• 50 samples for Monte Carlo Dropout sampling at test time

The following table outlines the testing results from Epistemic Tiramisu.

The following box plot describes the mean and standard deviation in wall clock time execution of different segmentation models performing inference on images of size 352 x 480 pixels.

The following box plot describes the mean and standard deviation in wall clock time execution of different Bayesian segmentation models performing inference on images of size 352 x 480 pixels. Note that in this case, inference is probabilistic due to the test time dropout and Monte Carlo simulation over 50 network samples.