Abstract by Michael Whitney
Deep Computer Vision Parameter Tuning
Many algorithms in the field of computer vision contain parameters that are either manually tuned or statically set. These parameters--which greatly impact the output--tend to rely heavily on the input image. Deep learning is a technique that has seen great success in solving computer vision problems. The goal of this work will be to use deep learning to produce the optimal parameter settings for classical computer vision algorithms. It allows for automatic parameter tuning for classical computer vision algorithms on a per-image basis to create a more robust system. In particular, this work focuses on automatically tuning parameters of the interactive graph cut segmentation algorithm. This algorithm contains one specific parameter that needs to be tuned and that has been the subject of much research in the field of computer vision. Using a combination of classification and regression methods, we use deep learning to achieve results greater than statically setting the parameter to an empirically good value.