Trust Your Model: Iterative Label Improvement and Robust Training by Confidence Based Filtering and Dataset Partitioning