MetaPix: Domain transfer for semantic segmentation by meta pixel weighting

Training a deep neural model for semantic segmentation requires collecting large amount of pixel-level labeled data. To alleviate the data scarcity problem presented in real world, one could utilize synthetic whose label is easy to obtain. Previous work has shown that performance can be improved by training jointly with and examples proper weighting on Such was learned heuristic maximize similarity between examples. In our work, we instead learn meta-learning, i.e., learning should only minimizing loss target task. We achieve this gradient-on-gradient technique propagate back into parameters model. The experiments show method single meta module outperform complicated combination an adversarial feature alignment, reconstruction loss, plus hierarchical at pixel, region image levels.