基于掩码对比学习的序列点云模型局部特征增强方法

In recent years, sequential point cloud models have demonstrated substantial potential in the field of point cloud processing. Compared with traditional KNN-based neighborhood feature aggregation methods, these sequential models enhance computational efficiency while expanding the receptive field by mapping point cloud data into spatially adjacent sequences. However, the expansion of the receptive field typically compromises the model\'s ability to capture fine-grained local features, resulting in insufficient learning of detailed local information. To address this limitation, this paper proposes a masked contrastive learning-based approach for enhancing local features in sequential point cloud models. Specifically, we design a unified framework integrating masked contrastive learning with supervised point cloud segmentation, where contrastive learning tasks are employed to intensively explore local characteristics within point cloud data. To achieve effective collaboration between self-supervised contrastive learning and supervised segmentation tasks, the model incorporates a regularization strategy that maintains feature consistency while balancing the learning objectives of both tasks. Experimental evaluations on the ShapeNet Part and S3DIS datasets demonstrate that the proposed method significantly improves segmentation performance in complex scenarios. Across multiple evaluation metrics, our approach outperforms existing models, conclusively validating the effectiveness of the proposed methodology.