Color-Difference Correntropy Guided Convolution Network for Point Cloud Semantic Segmentation

With the development of data acquisition technology, RGB color information is widely collected to strengthen the 3D point cloud. To some extent, RGB color information contains the prior relation about the spatial position of objects. However, the existing point cloud segmentation networks based on deep learning do not pay much attention to it. To remedy the lack in this area, we propose a color-difference correntropy guided convolution network, which introduces correntropy to optimize the measurement of color-difference. Meanwhile, we select points in the local neighborhood via the color-difference guided module, and construct an ordered sequence of points with correlation information, which not only facilitates the feature extraction by directly applying the convolution but also fully studies the correlation between the color information and spatial position of the point cloud. Moreover, we fuse the sequence features extracted by convolution with the geometric features acquired by MLP to get new features with more abundant semantic information, thus improving the segmentation performance. On both indoor and outdoor datasets, the experimental results demonstrate the effectiveness and superiority of the proposed method by the comparison experiments and ablation experiments.