View Decomposition and Adversarial for Semantic Segmentation

The adversarial training strategy has been effectively validated because it maintains high-level contextual consistency. However, limited to the weak capability of a simple discriminator, it is irresponsible and unreasonable to identify one from the sample source at a time. We introduce a novel discriminator module called Multi-View Decomposition which transforms the discriminator role from general teacher to specific adversary. The proposed module separates single sample into a series of class inter-independent streams and extracts corresponding features from current mask. The key insight in the MVD module is that the final source decision can be aggregated from all available views rather than a harsh critic. Our experimental results demonstrate that the proposed module can improve performance on PASCAL VOC 2012 and PASCAL Context dataset further.