Threshold-Adaptive Unsupervised Focal Loss for Domain Adaptation of Semantic Segmentation

Semantic segmentation is an important task for intelligent vehicles to understand the environment. Current deep learning based methods require large amounts of labeled data training. Manual annotation expensive, while simulators can provide accurate annotations. However, performance semantic model trained with synthetic datasets will significantly degenerate in actual scenes. Unsupervised domain adaptation (UDA) used reduce gap and improve on target domain. Existing adversarial-based self-training usually involve complex training procedures, entropy-based have recently received attention their simplicity effectiveness. UDA problems that they barely optimize hard samples lack explicit connection between source domains. In this paper, we propose a novel two-stage method segmentation. stage one, design threshold-adaptative unsupervised focal loss regularize prediction It first introduces into segmentation, helping avoiding generating unreliable pseudo-labels two, employ cross-domain image mixing (CIM) bridge knowledge two domains incorporate long-tail class pasting alleviate imbalance problem. Extensive experiments synthetic-to-real cross-city benchmarks demonstrate effectiveness our method. achieves state-of-the-art using DeepLabV2, as well competitive lightweight BiSeNet great advantages inference time.