Adapting Pre-Trained Vision Models for Novel Instance Detection and Segmentation