Cross-Modal Contrastive Learning for Domain Adaptation in 3D Semantic Segmentation

Domain adaptation for 3D point cloud has attracted a lot of interest since it can avoid the time-consuming labeling process data to some extent. A recent work named xMUDA leveraged multi-modal domain task semantic segmentation by mimicking predictions between 2D and modalities, outperformed previous single modality methods only using clouds. Based on it, in this paper, we propose novel cross-modal contrastive learning scheme further improve effects. By employing constraints from correspondences pixel features features, our method not facilitates interaction two different but also boosts feature representations both labeled source unlabeled target domain. Meanwhile, sufficiently utilize context information through learning, introduce neighborhood aggregation module enhance features. The employs attention aggregate nearby pixels image, which relieves mismatching arising projecting relative sparse dense image pixels. We evaluate three unsupervised scenarios, including country-to-country, day-to-night, dataset-to-dataset. Experimental results show that approach outperforms existing methods, demonstrates effectiveness proposed method.