Improvements to Image Reconstruction-Based Performance Prediction for Semantic Segmentation in Highly Automated Driving