HODOR: High-level Object Descriptors for Object Re-segmentation in Video Learned from Static Images