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

PROJECT TITLE :

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

ABSTRACT:

Significant headway has been made in semantic segmentation thanks to the quickening pace of development of convolutional neural networks, also known as CNNs. Such Deep Learning approaches require large scale real-world datasets with pixel-level annotations, despite the fact that they have been very successful. However, due to the fact that pixel-level labeling of semantics is a very labor-intensive process, a lot of researchers are turning to the use of synthetic data with free annotations. The segmentation model that was trained with synthetic images typically has poor performance when applied to real-world datasets because of the obvious gap in domain knowledge. Recent years have seen an increase in the amount of research attention paid to unsupervised domain adaptation (UDA), which is used for semantic segmentation. This research aims to reduce the domain discrepancy. Existing methods in this scope either simply align features or the outputs across the source and target domains or have to deal with the complex Image Processing and post-processing problems. Neither of these options is ideal. Confidence-and-Refinement Adaptation Model (CRAM) is a novel multi-level UDA model that we propose in this work. CRAM consists of two modules: a confidence-aware entropy alignment (CEA) module and a style feature alignment (SFA) module. The adaptation is carried out locally by means of adversarial learning in the output space when using CEA. This causes the segmentation model to pay attention to the predictions that have a high level of confidence. Additionally, the SFA module is applied in order to minimize the appearance gap between different domains. This is done in order to improve the model transfer in the shallow feature space. Experiments conducted on two difficult UDA benchmarks known as "GTA5-to-Cityscapes" and "SYNTHIA-to-Cityscapes" show that CRAM is effective. We achieve performance that is on par with the most recent works that are considered to be state-of-the-art while also offering advantages in terms of simplicity and convergence speed.

Did you like this research project?

To get this research project Guidelines, Training and Code... Click Here