Confidence-and-Refinement Adaptation Model for Cross-Domain Semantic Segmentation

With the rapid development of convolutional neural networks (CNNs), significant progress has been achieved in semantic segmentation. Despite great success, such deep learning approaches require large scale real-world datasets with pixel-level annotations. However, considering that labeling semantics is extremely laborious, many researchers turn to utilize synthetic data free But due clear domain gap, segmentation model trained images tends perform poorly on datasets. Unsupervised adaptation (UDA) for recently gains an increasing research attention, which aims at alleviating discrepancy. Existing methods this scope either simply align features or outputs across source and target domains have deal complex image processing post-processing problems. In work, we propose a novel multi-level UDA named Confidence-and-Refinement Adaptation Model (CRAM), contains confidence-aware entropy alignment (CEA) module style feature (SFA) module. Through CEA, done locally via adversarial output space, making pay attention high-confident predictions. Furthermore, enhance transfer shallow SFA applied minimize appearance gap domains. Experiments two challenging benchmarks “GTA5-to-Cityscapes” “SYNTHIA-to-Cityscapes” demonstrate effectiveness CRAM. We achieve comparable performance existing state-of-the-art works advantages simplicity convergence speed.