A Study on the Vulnerability of Semantic Segmentation Model to Data Transformation

초록

With the advancement of autonomous driving technology, the importance of semantic segmentation has markedly increased,while the amount of datasets needed for training has been limited. Accordingly, there has been a growing effort to increasedatasets using data augmentation techniques to train semantic segmentation models. However, the distributional gap betweenaugmented and real data can lead to performance limitations when models trained on real data are applied to augmented data. Therefore, this paper constructs new datasets by applying proposed data transformations on real-world datasets. Additionally, weevaluate the impact of these transformations on semantic segmentation models trained on real datasets. Results show that semanticsegmentation models are vulnerable to distortions in color information and object characteristics in transformed datasets. Furthermore, the vision transformer based model is less sensitive to distribution changes and shows greater segmentationperformance compared to fully convolutional network based models.

키워드