Enhancing Point Cloud Semantic Segmentation with Curriculum Learning for Occlusion Handling

Point cloud semantic segmentation is a critical task in 3D computer vision, particularly in applications such as autonomous robot. However, occlusions present a significant challenge, leading to incomplete data and reduced segmentation accuracy. In this paper, we propose a novel method that combines artificial occlusion with curriculum learning to enhance the robustness of segmentation models. Using the SPVCNN[2] model and the SemanticKITTI[3] dataset, we demonstrate that our approach significantly improves performance. Our curriculum learning strategy gradually increases the intensity and frequency of occlusions during training, enabling the model to better handle occluded regions. Experimental results show that our method achieves a improvement over the baseline method. These findings underscore the effectiveness of our approach in improving the accuracy and robustness of point cloud semantic segmentation.