UC Merced

Thanks to the development of deep neural networks, a number of computer vision tasks have achieved great success. However, the focus has been mostly limited to benchmarks with regular scenes in a supervised training fashion. A deep learning model trained with perfect and ideal benchmark datasets can have difficulty when applied to real-world scenes where the data are captured under different settings, for example. This indicates the model has poor generalization capability. Problems also occur when a benchmark model is applied to a different real-world application than it was designed for and where the input data varies. Therefore, this dissertation seeks to improve model generalization and adaptation for the computer vision problem of semantic segmentation particularly for real-world applications.