Evaluating Robustness of Neural Networks on Rotationally Disrupted Datasets for Semantic Segmentation

Here in this this study, I looked into the robustness of neural networks, specifically ICNet and U-Net, when trained and tested on datasets with rotationally disrupted distributions. My investigation involves deliberately partitioning test and training sets based on object angles from a set of data, diverging from the standard practice of random splitting. This approach allows us to assess how well these models generalize to unseen perspectives with limited training data. I utilize the VISAPP dataset, adjusting the models to handle binary and multi-class segmentation tasks. The evaluation is conducted using the Dice coefficient, considering the inherent challenges of multi-class generalization. My findings point towards the models’ exhibition of varying degrees of robustness, with performance significantly affected by the rotational split of the data. Interestingly, the most substantial drop in segmentation accuracy occurs when the training set contains extreme angles and the test set includes intermediate angles. These findings underscore the importance of dataset configuration in training deep learning models for semantic segmentation, particularly when data generalization across different perspectives is required.