Semantic Segmentation of buildings present in satellite images using encoder-decoder like convolutional neural networks is being achieved with relatively high pixel-wise metric scores. In this paper, we aim to exploit the power of fully convolutional neural networks for an instance segmentation task using extra added classes to the output along with the watershed processing technique to leverage better object-wise metric results. We also show that CutMix mixed data augmentations and the One-Cycle learning rate policy are greater regularization methods to achieve a better fit on the training data and increase performance. Furthermore, Mixed Precision Training provided more flexibility to experiment with bigger networks and batches while maintaining stability and convergence during training. We compare and show the effect of these additional changes throughout our whole pipeline to finally provide a set a tuned hyper-parameters that are proven to perform better.
View on arXiv