Lookahead adversarial learning for near real-time semantic segmentation

Semantic segmentation is one of the most fundamental problems in computer vision with significant impact on a wide variety applications. Adversarial learning shown to be an effective approach for improving semantic quality by enforcing higher-level pixel correlations and structural information. However, state-of-the-art models cannot easily plugged into adversarial setting because they are not designed accommodate convergence stability issues networks. We bridge this gap building conditional network model (DeepLabv3+) at its core. To battle issues, we introduce novel lookahead (LoAd) embedded label map aggregation module. focus that run fast inference near real-time field Through extensive experimentation, demonstrate proposed solution can alleviate divergence results considerable performance improvements (+5% some classes) baseline three standard datasets.